Sepsis, the systemic inflammatory response to infection, is a major public health concern.[1] Worldwide, sepsis affects millions of hospitalized patients each year.[2] In the United States, it is the single most expensive cause of hospitalization.[3, 4, 5, 6] Multiple studies suggest that sepsis hospitalizations are also increasing in frequency.[3, 6, 7, 8, 9, 10]

Improved sepsis care has dramatically reduced in‐hospital mortality.[11, 12, 13] However, the result is a growing number of sepsis survivors discharged with new disability.[1, 9, 14, 15, 16] Despite being a common cause of hospitalization, little is known about how to improve postsepsis care.[15, 17, 18, 19] This contrasts with other, often less common, hospital conditions for which many studies evaluating readmission and postdischarge care are available.[20, 21, 22, 23] Identifying the factors contributing to high utilization could lend critical insight to designing interventions that improve long‐term sepsis outcomes.[24]

We conducted a retrospective study of sepsis patients discharged in 2010 at Kaiser Permanente Northern California (KPNC) to describe their posthospital trajectories. In this diverse community‐hospitalbased population, we sought to identify the patient‐level factors that impact the posthospital healthcare utilization of sepsis survivors.

METHODS

This study was approved by the KPNC institutional review board.

RESULTS

One‐Year Survival

A total of 5479 (86.4%) patients survived their index sepsis hospitalization. Overall survival after living discharge was 90.5% (range, 89.6%91.2%) at 30 days and 71.3% (range, 70.1%72.5%) at 1 year. However, postsepsis survival was strongly modified by age (Figure 1). For example, 1‐year survival was 94.1% (range, 91.2%96.0%) for <45 year olds and 54.4% (range, 51.5%57.2%) for 85 year olds (P<0.01). Survival was also modified by predicted mortality, however, not by ICU admission during index hospitalization (P=0.18) (see Supporting Appendix, Figure 1, in the online version of this article).

Figure 1

Healthcare Utilization

The unadjusted difference between pre‐ and postsepsis healthcare utilization among survivors was statistically significant for most categories but of modest clinical significance (see Supporting Appendix, Table 4, in the online version of this article). For example, the mean number of presepsis hospitalizations was 0.9 (1.4) compared to 1.0 (1.5) postsepsis (P<0.01). After adjusting for postsepsis living days, the difference in utilization was more pronounced (Figure 2). Overall, there was roughly a 3‐fold increase in the mean percentage of living days spent in facility‐based care between patients' pre‐ and postsepsis phases (5.3% vs 15.0%, P<0.01). Again, the difference was strongly modified by age. For patients aged <45 years, the difference was not statistically significant (2.4% vs 2.9%, P=0.32), whereas for those aged 65 years, it was highly significant (6.2% vs 18.5%, P<0.01).

Figure 2

Factors associated with early readmission included severity of illness, comorbid disease burden, index hospital length of stay, and intensive care (Table 3). However, the dominant factor explaining variation in the risk of early readmission was patients' prior comorbid disease burden (73.9%), followed by acute severity of illness (12.4%), total hospital length of stay (6.6%), and the need for intensive care (5.2%). Severity of illness and age were also significantly associated with higher odds of high postsepsis utilization; however, the dominant factor contributing to this risk was a history of high presepsis utilization (64.2%).

DISCUSSION

In this population‐based study in a community healthcare system, the impact of sepsis extended well beyond the initial hospitalization. One in 6 sepsis survivors was readmitted within 30 days, and roughly half were readmitted within 1 year. Fewer than half of rehospitalizations were for sepsis. Patients had a 3‐fold increase in the percentage of living days spent in hospitals or care facilities after sepsis hospitalization. Although age and acute severity of illness strongly modified healthcare utilization and mortality after sepsis, the dominant factors contributing to early readmission and high utilization ratescomorbid disease burden and presepsis healthcare utilizationwere present prior to hospitalization.

Sepsis is the single most expensive cause of US hospitalizations.[3, 4, 5] Despite its prevalence, there are little contemporary data identifying factors that impact healthcare utilization among sepsis survivors.[9, 16, 17, 19, 24, 36, 37] Recently, Prescott and others found that in Medicare beneficiaries, following severe sepsis, healthcare utilization was markedly increased.[17] More than one‐quarter of survivors were readmitted within 30 days, and 63.8% were readmitted within a year. Severe sepsis survivors also spent an average of 26% of their living days in a healthcare facility, a nearly 4‐fold increase compared to their presepsis phase. The current study included a population with a broader age and severity range; however, in a similar subgroup of patients, for those aged 65 years within the highest predicted mortality quartile, the frequency of readmission was similar. These findings are concordant with those from prior studies.[17, 19, 36, 37]

Among sepsis survivors, most readmissions were not for sepsis or infectious diagnoses, which is a novel finding with implications for designing approaches to reduce rehospitalization. The pattern in sepsis is similar to that seen in other common and costly hospital conditions.[17, 20, 23, 38, 39, 40] For example, between 18% and 25% of Medicare beneficiaries hospitalized for heart failure, acute myocardial infarction, or pneumonia were readmitted within 30 days; fewer than one‐third had the same diagnosis.[20] The timing of readmission in our sepsis cohort was also similar to that seen in other conditions.[20] For example, the median time of early readmission in this study was 11 days; it was between 10 and 12 days for patients with heart failure, pneumonia, and myocardial infarction.[20]

Krumholz and others suggest that the pattern of early rehospitalization after common acute conditions reflects a posthospital syndromean acquired, transient period of vulnerabilitythat could be the byproduct of common hospital factors.[20, 41] Such universal impairments might result from new physical and neurocognitive disability, nutritional deficiency, and sleep deprivation or delirium, among others.[41] If this construct were also true in sepsis, it could have important implications on the design of postsepsis care. However, prior studies suggest that sepsis patients may be particularly vulnerable to the sequelae of hospitalization.[2, 42, 43, 44, 45]

Among Medicare beneficiaries, Iwashyna and others reported that hospitalizations for severe sepsis resulted in significant increases in physical limitations and moderate to severe cognitive impairment.[1, 14, 46] Encephalopathy, sleep deprivation, and delirium are also frequently seen in sepsis patients.[47, 48] Furthermore, sepsis patients frequently need intensive care, which is also associated with increased patient disability and injury.[16, 46, 49, 50] We found that severity of illness and the need for intensive care were both predictive of the need for early readmission following sepsis. We also confirmed the results of prior studies suggesting that sepsis outcomes are strongly modified by age.[16, 19, 43, 51]

However, we found that the dominant factors contributing to patients' health trajectories were conditions present prior to admission. This finding is in accord with prior suggestions that acute severity of illness only partially predicts patients facing adverse posthospital sequelae.[23, 41, 52] Among sepsis patients, prior work demonstrates that inadequate consideration for presepsis level of function and utilization can result in an overestimation of the impact of sepsis on postdischarge health.[52, 53] Further, we found that the need for intensive care was not independently associated with an increased risk of high postsepsis utilization after adjusting for illness severity, a finding also seen in prior studies.[17, 23, 38, 51]

Taken together, our findings might suggest that an optimal approach to posthospital care in sepsis should focus on treatment approaches that address disease‐specific problems within the much larger context of common hospital risks. However, further study is necessary to clearly define the mechanisms by which age, severity of illness, and intensive care affect subsequent healthcare utilization. Furthermore, sepsis patients are a heterogeneous population in terms of severity of illness, site and pathogen of infection, and underlying comorbidity whose posthospital course remains incompletely characterized, limiting our ability to draw strong inferences.

These results should be interpreted in light of the study's limitations. First, our cohort included patients with healthcare insurance within a community‐based healthcare system. Care within the KPNC system, which bears similarities with accountable care organizations, is enhanced through service integration and a comprehensive health information system. Although prior studies suggest that these characteristics result in improved population‐based care, it is unclear whether there is a similar impact in hospital‐based conditions such as sepsis.[54, 55] Furthermore, care within an integrated system may impact posthospital utilization patterns and could limit generalizability. However, prior studies demonstrate the similarity of KPNC members to other patients in the same region in terms of age, socioeconomics, overall health behaviors, and racial/ethnic diversity.[56] Second, our study did not characterize organ dysfunction based on diagnosis coding, a common feature of sepsis studies that lack detailed physiologic severity data.[4, 5, 6, 8, 26] Instead, we focused on using granular laboratory and vital signs data to ensure accurate risk adjustment using a validated system developed in >400,000 hospitalizations.[30] Although this method may hamper comparisons with existing studies, traditional methods of grading severity by diagnosis codes can be vulnerable to biases resulting in wide variability.[10, 23, 26, 57, 58] Nonetheless, it is likely that characterizing preexisting and acute organ dysfunction will improve risk stratification in the heterogeneous sepsis population. Third, this study did not include data regarding patients' functional status, which has been shown to strongly predict patient outcomes following hospitalization. Fourth, this study did not address the cost of care following sepsis hospitalizations.[19, 59] Finally, our study excluded patients with incomplete utilization records, a choice designed to avoid the spurious inferences that can result from such comparisons.[53]

In summary, we found that sepsis exacted a considerable toll on patients in the hospital and in the year following discharge. Sepsis patients were frequently rehospitalized within a month of discharge, and on average had a 3‐fold increase in their subsequent time spent in healthcare facilities. Although age, severity of illness, and the need for ICU care impacted postsepsis utilization, the dominant contributing factorscomorbid disease burden or presepsis utilizationwere present prior to sepsis hospitalization. Early readmission patterns in sepsis appeared similar to those seen in other important hospital conditions, suggesting a role for shared posthospital, rather than just postsepsis, care approaches.

Sepsis, the systemic inflammatory response to infection, is a major public health concern.[1] Worldwide, sepsis affects millions of hospitalized patients each year.[2] In the United States, it is the single most expensive cause of hospitalization.[3, 4, 5, 6] Multiple studies suggest that sepsis hospitalizations are also increasing in frequency.[3, 6, 7, 8, 9, 10]

Improved sepsis care has dramatically reduced in‐hospital mortality.[11, 12, 13] However, the result is a growing number of sepsis survivors discharged with new disability.[1, 9, 14, 15, 16] Despite being a common cause of hospitalization, little is known about how to improve postsepsis care.[15, 17, 18, 19] This contrasts with other, often less common, hospital conditions for which many studies evaluating readmission and postdischarge care are available.[20, 21, 22, 23] Identifying the factors contributing to high utilization could lend critical insight to designing interventions that improve long‐term sepsis outcomes.[24]

We conducted a retrospective study of sepsis patients discharged in 2010 at Kaiser Permanente Northern California (KPNC) to describe their posthospital trajectories. In this diverse community‐hospitalbased population, we sought to identify the patient‐level factors that impact the posthospital healthcare utilization of sepsis survivors.

METHODS

This study was approved by the KPNC institutional review board.

RESULTS

One‐Year Survival

A total of 5479 (86.4%) patients survived their index sepsis hospitalization. Overall survival after living discharge was 90.5% (range, 89.6%91.2%) at 30 days and 71.3% (range, 70.1%72.5%) at 1 year. However, postsepsis survival was strongly modified by age (Figure 1). For example, 1‐year survival was 94.1% (range, 91.2%96.0%) for <45 year olds and 54.4% (range, 51.5%57.2%) for 85 year olds (P<0.01). Survival was also modified by predicted mortality, however, not by ICU admission during index hospitalization (P=0.18) (see Supporting Appendix, Figure 1, in the online version of this article).

Figure 1

Healthcare Utilization

The unadjusted difference between pre‐ and postsepsis healthcare utilization among survivors was statistically significant for most categories but of modest clinical significance (see Supporting Appendix, Table 4, in the online version of this article). For example, the mean number of presepsis hospitalizations was 0.9 (1.4) compared to 1.0 (1.5) postsepsis (P<0.01). After adjusting for postsepsis living days, the difference in utilization was more pronounced (Figure 2). Overall, there was roughly a 3‐fold increase in the mean percentage of living days spent in facility‐based care between patients' pre‐ and postsepsis phases (5.3% vs 15.0%, P<0.01). Again, the difference was strongly modified by age. For patients aged <45 years, the difference was not statistically significant (2.4% vs 2.9%, P=0.32), whereas for those aged 65 years, it was highly significant (6.2% vs 18.5%, P<0.01).

Figure 2

Factors associated with early readmission included severity of illness, comorbid disease burden, index hospital length of stay, and intensive care (Table 3). However, the dominant factor explaining variation in the risk of early readmission was patients' prior comorbid disease burden (73.9%), followed by acute severity of illness (12.4%), total hospital length of stay (6.6%), and the need for intensive care (5.2%). Severity of illness and age were also significantly associated with higher odds of high postsepsis utilization; however, the dominant factor contributing to this risk was a history of high presepsis utilization (64.2%).

DISCUSSION

In this population‐based study in a community healthcare system, the impact of sepsis extended well beyond the initial hospitalization. One in 6 sepsis survivors was readmitted within 30 days, and roughly half were readmitted within 1 year. Fewer than half of rehospitalizations were for sepsis. Patients had a 3‐fold increase in the percentage of living days spent in hospitals or care facilities after sepsis hospitalization. Although age and acute severity of illness strongly modified healthcare utilization and mortality after sepsis, the dominant factors contributing to early readmission and high utilization ratescomorbid disease burden and presepsis healthcare utilizationwere present prior to hospitalization.

Sepsis is the single most expensive cause of US hospitalizations.[3, 4, 5] Despite its prevalence, there are little contemporary data identifying factors that impact healthcare utilization among sepsis survivors.[9, 16, 17, 19, 24, 36, 37] Recently, Prescott and others found that in Medicare beneficiaries, following severe sepsis, healthcare utilization was markedly increased.[17] More than one‐quarter of survivors were readmitted within 30 days, and 63.8% were readmitted within a year. Severe sepsis survivors also spent an average of 26% of their living days in a healthcare facility, a nearly 4‐fold increase compared to their presepsis phase. The current study included a population with a broader age and severity range; however, in a similar subgroup of patients, for those aged 65 years within the highest predicted mortality quartile, the frequency of readmission was similar. These findings are concordant with those from prior studies.[17, 19, 36, 37]

Among sepsis survivors, most readmissions were not for sepsis or infectious diagnoses, which is a novel finding with implications for designing approaches to reduce rehospitalization. The pattern in sepsis is similar to that seen in other common and costly hospital conditions.[17, 20, 23, 38, 39, 40] For example, between 18% and 25% of Medicare beneficiaries hospitalized for heart failure, acute myocardial infarction, or pneumonia were readmitted within 30 days; fewer than one‐third had the same diagnosis.[20] The timing of readmission in our sepsis cohort was also similar to that seen in other conditions.[20] For example, the median time of early readmission in this study was 11 days; it was between 10 and 12 days for patients with heart failure, pneumonia, and myocardial infarction.[20]

Krumholz and others suggest that the pattern of early rehospitalization after common acute conditions reflects a posthospital syndromean acquired, transient period of vulnerabilitythat could be the byproduct of common hospital factors.[20, 41] Such universal impairments might result from new physical and neurocognitive disability, nutritional deficiency, and sleep deprivation or delirium, among others.[41] If this construct were also true in sepsis, it could have important implications on the design of postsepsis care. However, prior studies suggest that sepsis patients may be particularly vulnerable to the sequelae of hospitalization.[2, 42, 43, 44, 45]

Among Medicare beneficiaries, Iwashyna and others reported that hospitalizations for severe sepsis resulted in significant increases in physical limitations and moderate to severe cognitive impairment.[1, 14, 46] Encephalopathy, sleep deprivation, and delirium are also frequently seen in sepsis patients.[47, 48] Furthermore, sepsis patients frequently need intensive care, which is also associated with increased patient disability and injury.[16, 46, 49, 50] We found that severity of illness and the need for intensive care were both predictive of the need for early readmission following sepsis. We also confirmed the results of prior studies suggesting that sepsis outcomes are strongly modified by age.[16, 19, 43, 51]

However, we found that the dominant factors contributing to patients' health trajectories were conditions present prior to admission. This finding is in accord with prior suggestions that acute severity of illness only partially predicts patients facing adverse posthospital sequelae.[23, 41, 52] Among sepsis patients, prior work demonstrates that inadequate consideration for presepsis level of function and utilization can result in an overestimation of the impact of sepsis on postdischarge health.[52, 53] Further, we found that the need for intensive care was not independently associated with an increased risk of high postsepsis utilization after adjusting for illness severity, a finding also seen in prior studies.[17, 23, 38, 51]

Taken together, our findings might suggest that an optimal approach to posthospital care in sepsis should focus on treatment approaches that address disease‐specific problems within the much larger context of common hospital risks. However, further study is necessary to clearly define the mechanisms by which age, severity of illness, and intensive care affect subsequent healthcare utilization. Furthermore, sepsis patients are a heterogeneous population in terms of severity of illness, site and pathogen of infection, and underlying comorbidity whose posthospital course remains incompletely characterized, limiting our ability to draw strong inferences.

These results should be interpreted in light of the study's limitations. First, our cohort included patients with healthcare insurance within a community‐based healthcare system. Care within the KPNC system, which bears similarities with accountable care organizations, is enhanced through service integration and a comprehensive health information system. Although prior studies suggest that these characteristics result in improved population‐based care, it is unclear whether there is a similar impact in hospital‐based conditions such as sepsis.[54, 55] Furthermore, care within an integrated system may impact posthospital utilization patterns and could limit generalizability. However, prior studies demonstrate the similarity of KPNC members to other patients in the same region in terms of age, socioeconomics, overall health behaviors, and racial/ethnic diversity.[56] Second, our study did not characterize organ dysfunction based on diagnosis coding, a common feature of sepsis studies that lack detailed physiologic severity data.[4, 5, 6, 8, 26] Instead, we focused on using granular laboratory and vital signs data to ensure accurate risk adjustment using a validated system developed in >400,000 hospitalizations.[30] Although this method may hamper comparisons with existing studies, traditional methods of grading severity by diagnosis codes can be vulnerable to biases resulting in wide variability.[10, 23, 26, 57, 58] Nonetheless, it is likely that characterizing preexisting and acute organ dysfunction will improve risk stratification in the heterogeneous sepsis population. Third, this study did not include data regarding patients' functional status, which has been shown to strongly predict patient outcomes following hospitalization. Fourth, this study did not address the cost of care following sepsis hospitalizations.[19, 59] Finally, our study excluded patients with incomplete utilization records, a choice designed to avoid the spurious inferences that can result from such comparisons.[53]

In summary, we found that sepsis exacted a considerable toll on patients in the hospital and in the year following discharge. Sepsis patients were frequently rehospitalized within a month of discharge, and on average had a 3‐fold increase in their subsequent time spent in healthcare facilities. Although age, severity of illness, and the need for ICU care impacted postsepsis utilization, the dominant contributing factorscomorbid disease burden or presepsis utilizationwere present prior to sepsis hospitalization. Early readmission patterns in sepsis appeared similar to those seen in other important hospital conditions, suggesting a role for shared posthospital, rather than just postsepsis, care approaches.

Sepsis, the systemic inflammatory response to infection, is a major public health concern.[1] Worldwide, sepsis affects millions of hospitalized patients each year.[2] In the United States, it is the single most expensive cause of hospitalization.[3, 4, 5, 6] Multiple studies suggest that sepsis hospitalizations are also increasing in frequency.[3, 6, 7, 8, 9, 10]

Improved sepsis care has dramatically reduced in‐hospital mortality.[11, 12, 13] However, the result is a growing number of sepsis survivors discharged with new disability.[1, 9, 14, 15, 16] Despite being a common cause of hospitalization, little is known about how to improve postsepsis care.[15, 17, 18, 19] This contrasts with other, often less common, hospital conditions for which many studies evaluating readmission and postdischarge care are available.[20, 21, 22, 23] Identifying the factors contributing to high utilization could lend critical insight to designing interventions that improve long‐term sepsis outcomes.[24]

We conducted a retrospective study of sepsis patients discharged in 2010 at Kaiser Permanente Northern California (KPNC) to describe their posthospital trajectories. In this diverse community‐hospitalbased population, we sought to identify the patient‐level factors that impact the posthospital healthcare utilization of sepsis survivors.

METHODS

This study was approved by the KPNC institutional review board.

RESULTS

One‐Year Survival

A total of 5479 (86.4%) patients survived their index sepsis hospitalization. Overall survival after living discharge was 90.5% (range, 89.6%91.2%) at 30 days and 71.3% (range, 70.1%72.5%) at 1 year. However, postsepsis survival was strongly modified by age (Figure 1). For example, 1‐year survival was 94.1% (range, 91.2%96.0%) for <45 year olds and 54.4% (range, 51.5%57.2%) for 85 year olds (P<0.01). Survival was also modified by predicted mortality, however, not by ICU admission during index hospitalization (P=0.18) (see Supporting Appendix, Figure 1, in the online version of this article).

Figure 1

Healthcare Utilization

The unadjusted difference between pre‐ and postsepsis healthcare utilization among survivors was statistically significant for most categories but of modest clinical significance (see Supporting Appendix, Table 4, in the online version of this article). For example, the mean number of presepsis hospitalizations was 0.9 (1.4) compared to 1.0 (1.5) postsepsis (P<0.01). After adjusting for postsepsis living days, the difference in utilization was more pronounced (Figure 2). Overall, there was roughly a 3‐fold increase in the mean percentage of living days spent in facility‐based care between patients' pre‐ and postsepsis phases (5.3% vs 15.0%, P<0.01). Again, the difference was strongly modified by age. For patients aged <45 years, the difference was not statistically significant (2.4% vs 2.9%, P=0.32), whereas for those aged 65 years, it was highly significant (6.2% vs 18.5%, P<0.01).

Figure 2

Factors associated with early readmission included severity of illness, comorbid disease burden, index hospital length of stay, and intensive care (Table 3). However, the dominant factor explaining variation in the risk of early readmission was patients' prior comorbid disease burden (73.9%), followed by acute severity of illness (12.4%), total hospital length of stay (6.6%), and the need for intensive care (5.2%). Severity of illness and age were also significantly associated with higher odds of high postsepsis utilization; however, the dominant factor contributing to this risk was a history of high presepsis utilization (64.2%).

DISCUSSION

In this population‐based study in a community healthcare system, the impact of sepsis extended well beyond the initial hospitalization. One in 6 sepsis survivors was readmitted within 30 days, and roughly half were readmitted within 1 year. Fewer than half of rehospitalizations were for sepsis. Patients had a 3‐fold increase in the percentage of living days spent in hospitals or care facilities after sepsis hospitalization. Although age and acute severity of illness strongly modified healthcare utilization and mortality after sepsis, the dominant factors contributing to early readmission and high utilization ratescomorbid disease burden and presepsis healthcare utilizationwere present prior to hospitalization.

Sepsis is the single most expensive cause of US hospitalizations.[3, 4, 5] Despite its prevalence, there are little contemporary data identifying factors that impact healthcare utilization among sepsis survivors.[9, 16, 17, 19, 24, 36, 37] Recently, Prescott and others found that in Medicare beneficiaries, following severe sepsis, healthcare utilization was markedly increased.[17] More than one‐quarter of survivors were readmitted within 30 days, and 63.8% were readmitted within a year. Severe sepsis survivors also spent an average of 26% of their living days in a healthcare facility, a nearly 4‐fold increase compared to their presepsis phase. The current study included a population with a broader age and severity range; however, in a similar subgroup of patients, for those aged 65 years within the highest predicted mortality quartile, the frequency of readmission was similar. These findings are concordant with those from prior studies.[17, 19, 36, 37]

Among sepsis survivors, most readmissions were not for sepsis or infectious diagnoses, which is a novel finding with implications for designing approaches to reduce rehospitalization. The pattern in sepsis is similar to that seen in other common and costly hospital conditions.[17, 20, 23, 38, 39, 40] For example, between 18% and 25% of Medicare beneficiaries hospitalized for heart failure, acute myocardial infarction, or pneumonia were readmitted within 30 days; fewer than one‐third had the same diagnosis.[20] The timing of readmission in our sepsis cohort was also similar to that seen in other conditions.[20] For example, the median time of early readmission in this study was 11 days; it was between 10 and 12 days for patients with heart failure, pneumonia, and myocardial infarction.[20]

Krumholz and others suggest that the pattern of early rehospitalization after common acute conditions reflects a posthospital syndromean acquired, transient period of vulnerabilitythat could be the byproduct of common hospital factors.[20, 41] Such universal impairments might result from new physical and neurocognitive disability, nutritional deficiency, and sleep deprivation or delirium, among others.[41] If this construct were also true in sepsis, it could have important implications on the design of postsepsis care. However, prior studies suggest that sepsis patients may be particularly vulnerable to the sequelae of hospitalization.[2, 42, 43, 44, 45]

Among Medicare beneficiaries, Iwashyna and others reported that hospitalizations for severe sepsis resulted in significant increases in physical limitations and moderate to severe cognitive impairment.[1, 14, 46] Encephalopathy, sleep deprivation, and delirium are also frequently seen in sepsis patients.[47, 48] Furthermore, sepsis patients frequently need intensive care, which is also associated with increased patient disability and injury.[16, 46, 49, 50] We found that severity of illness and the need for intensive care were both predictive of the need for early readmission following sepsis. We also confirmed the results of prior studies suggesting that sepsis outcomes are strongly modified by age.[16, 19, 43, 51]

However, we found that the dominant factors contributing to patients' health trajectories were conditions present prior to admission. This finding is in accord with prior suggestions that acute severity of illness only partially predicts patients facing adverse posthospital sequelae.[23, 41, 52] Among sepsis patients, prior work demonstrates that inadequate consideration for presepsis level of function and utilization can result in an overestimation of the impact of sepsis on postdischarge health.[52, 53] Further, we found that the need for intensive care was not independently associated with an increased risk of high postsepsis utilization after adjusting for illness severity, a finding also seen in prior studies.[17, 23, 38, 51]

Taken together, our findings might suggest that an optimal approach to posthospital care in sepsis should focus on treatment approaches that address disease‐specific problems within the much larger context of common hospital risks. However, further study is necessary to clearly define the mechanisms by which age, severity of illness, and intensive care affect subsequent healthcare utilization. Furthermore, sepsis patients are a heterogeneous population in terms of severity of illness, site and pathogen of infection, and underlying comorbidity whose posthospital course remains incompletely characterized, limiting our ability to draw strong inferences.

These results should be interpreted in light of the study's limitations. First, our cohort included patients with healthcare insurance within a community‐based healthcare system. Care within the KPNC system, which bears similarities with accountable care organizations, is enhanced through service integration and a comprehensive health information system. Although prior studies suggest that these characteristics result in improved population‐based care, it is unclear whether there is a similar impact in hospital‐based conditions such as sepsis.[54, 55] Furthermore, care within an integrated system may impact posthospital utilization patterns and could limit generalizability. However, prior studies demonstrate the similarity of KPNC members to other patients in the same region in terms of age, socioeconomics, overall health behaviors, and racial/ethnic diversity.[56] Second, our study did not characterize organ dysfunction based on diagnosis coding, a common feature of sepsis studies that lack detailed physiologic severity data.[4, 5, 6, 8, 26] Instead, we focused on using granular laboratory and vital signs data to ensure accurate risk adjustment using a validated system developed in >400,000 hospitalizations.[30] Although this method may hamper comparisons with existing studies, traditional methods of grading severity by diagnosis codes can be vulnerable to biases resulting in wide variability.[10, 23, 26, 57, 58] Nonetheless, it is likely that characterizing preexisting and acute organ dysfunction will improve risk stratification in the heterogeneous sepsis population. Third, this study did not include data regarding patients' functional status, which has been shown to strongly predict patient outcomes following hospitalization. Fourth, this study did not address the cost of care following sepsis hospitalizations.[19, 59] Finally, our study excluded patients with incomplete utilization records, a choice designed to avoid the spurious inferences that can result from such comparisons.[53]

In summary, we found that sepsis exacted a considerable toll on patients in the hospital and in the year following discharge. Sepsis patients were frequently rehospitalized within a month of discharge, and on average had a 3‐fold increase in their subsequent time spent in healthcare facilities. Although age, severity of illness, and the need for ICU care impacted postsepsis utilization, the dominant contributing factorscomorbid disease burden or presepsis utilizationwere present prior to sepsis hospitalization. Early readmission patterns in sepsis appeared similar to those seen in other important hospital conditions, suggesting a role for shared posthospital, rather than just postsepsis, care approaches.

Clostridium difficile infection (CDI) is a serious and costly condition whose volume in US hospitals has doubled over the last decade.[1, 2, 3] Along with this rise in incidence, its severity has also increased. Although in the United States there has been a doubling in age‐adjusted case fatality, in the same time period Canadian studies reported a high and increasing CDI‐associated case fatality in the setting of an outbreak of a novel epidemic hypervirulent strain BI/NAP1/027.[2, 4, 5, 6] The costs of CDI range widely ($2500 to $13,000 per hospitalization), with cumulative annual cost to the US healthcare system estimated at nearly $5 billion.[7, 8, 9]

One of the drivers of these clinical and economic outcomes is CDI recurrence (rCDI). In 2 recent randomized controlled trials, up to 25% of patients with an initial CDI (iCDI) episode developed rCDI.[10, 11] There are few data that quantify the impact of rCDI on quality of life and survival. However, patients often are readmitted to the hospital with rCDI, and physicians who treat patients with multiple episodes of rCDI can attest to the devastating toll it takes on the lives of the patients and their families (personal communications from numerous patients to E.R.D.).[12] Reducing the incidence of rCDI may significantly improve the course of this disease.

The advent of such new treatments as fidaxomicin aimed at rCDI is promising.[10, 11] However, evidence for its efficacy so far is limited to treatment‐naive iCDI patients, thus challenging clinicians to identify patients at high risk for rCDI at iCDI onset. To address this challenge, we set out to develop a bedside prediction model for rCDI based on the factors present and routinely available at the onset of iCDI.

METHODS

Study Design and Data Source

We conducted a retrospective single‐center cohort study to examine the factors present at the onset of iCDI that impact the incidence of rCDI among hospitalized patients. Patients were included in the study if they were adults (18 years) hospitalized at Barnes‐Jewish Hospital (BJH), St. Louis, Missouri, between January 1, 2003 and December 31, 2009, and who had a positive toxin assay for C difficile in the setting of unformed stools and no history of CDI in the previous 60 days (as defined by positive toxin assay). Patients were excluded if they either died during or were discharged to hospice from the iCDI hospitalization. Cases of iCDI were categorized according to published surveillance definitions as community onset‐healthcare facility associated (CO‐HCFA), healthcare facility onset, and community associated.[13] Notably, the CO‐HCFA category included surveillance definitions for both CO‐HCFA and indeterminate cases. We defined rCDI as a repeat positive toxin within 42 days following the end of iCDI treatment. This period of risk for rCDI was chosen because the current surveillance definition for rCDI is a new episode of CDI occurring within 8 weeks from the last episode of CDI, with the assumption the patient would receive 10 to 14 days of CDI treatment at the beginning of the 8‐week period.[14] Medical charts were reviewed for all readmissions during the recurrence risk period to identify patients diagnosed with rCDI by methods other than toxin assay. A study enrollment flow chart is shown in Figure 1.

Figure 1

Demographic and clinical data were derived from the BJH medical informatics databases and the BJH electronic medical records (see Supporting Appendix Table 1 in the online version of this article). Comorbidities were grouped using the Charlson‐Deyo categories.[15] All variables were limited to data that are consistent throughout a hospitalization (eg, race or age) or were present within 48 hours of iCDI (eg, medications).

RESULTS

Among the 4196 patients with iCDI enrolled in the study, 425 (10.1%) developed at least 1 recurrence within 42 days of the end of iCDI treatment (Table 1). Compared to patients without a recurrence, in univariate analysis those with an rCDI episode were older and had a greater comorbidity burden. In particular, diabetes mellitus (odds ratio 1.34; 95% confidence interval [CI], 1.08‐1.66) and cerebrovascular disease (odds ratio 1.47; 95% CI, 1.04‐2.08) were significantly more prevalent in the rCDI group. The index CDI episode for patients with rCDI was approximately twice as likely to fit the surveillance definition for CO‐HCFA than the index episode for those without a recurrence (odds ratio 2.24; 95% CI, 1.80‐2.79). Commensurately, patients with rCDI also had greater odds for experiencing multiple recent hospitalizations than those without rCDI. Neither type of CDI treatment (oral metronidazole vs oral vancomycin vs both), nor duration, was significantly associated with recurrence.

Seven factors present at the onset of iCDI were found to predict a recurrence in multivariable analysis (Table 2). Older age, CO‐HCFA status of iCDI, and 2 or more hospitalizations in the prior 60 days increased the risk of rCDI. Concomitant exposures to gastric acid suppressors, fluoroquinolone antibiotics, and high‐risk antibiotics were also significantly associated with a recurrence. Being in the intensive care unit (ICU) at the onset of iCDI was protective against rCDI in the multivariable model. This model had a C statistic of 0.642 and a Brier score of 0.089. After cross‐validation with 500 bootstrapping iterations, the model exhibited a moderately good fit (Figure 2). The prediction was particularly accurate in the lower risk ranges, with slight divergence in the risk strata over 20%. The validated model had a C statistic of 0.630 and Brier score of 0.089.

Figure 2

The sensitivity, specificity, and positive and negative predictive values of the model at various probability thresholds of rCDI are presented in Table 3. Thus, when the probability of rCDI was low, the model exhibited high sensitivity and low specificity. The situation was reversed as the probability of rCDI approached 30% (very low sensitivity and high specificity). The model's performance was optimal when the rCDI risk matched that in the current cohort, or 10.1%, with a sensitivity of 56% and specificity of 65%. However, when the rCDI risk dropped to 5%, the specificity dropped to below 30%. The sensitivity dropped to below 30% when rCDI risk rose to 15% (Table 3). Across the entire range of the probabilities tested, the negative predictive value of the model was persistently 90% or higher.

DISCUSSION

We have demonstrated that in a cohort of hospitalized patients with iCDI, 10% developed at least 1 episode of rCDI within 42 days of the end of iCDI treatment. The factors present at iCDI onset that predicted recurrence were age, CO‐HCFA CDI, prior hospitalization, high‐risk antibiotic and fluoroquinolone use, and gastric acid suppression. Although the model's performance was only moderate, its negative predictive value was 90% or higher across the entire range of rCDI probabilities tested. This means that the absence of this combination of risk factors in a patient with iCDI diminishes the probability of a rCDI episode to 10% or below, depending on the prior population risk for rCDI.

Prior investigators have developed prediction rules for rCDI. Hebert et al., using methodology similar to ours, constructed a model to predict the risk of rCDI among patients hospitalized with iCDI.[17] For example, the recurrence rate in their study was 23% compared to our 10%. This is likely due to the differing definitions of both iCDI and rCDI between the 2 studies. Although our definitions of hospital‐associated C difficile‐associated diarrhea (CDAD) conformed to the recommended surveillance definitions,[13] Hebert and colleagues used different definitions.[18] If this is so, the higher rate of rCDI in their study may have reflected these differences in surveillance definition, rather than the true prevalence of recurrent CDAD.

Several other studies have relied on either specialized laboratory tests alone or in combination with clinical factors. Stewart et al., in a small single‐center cohort study, reported the presence of the binary toxin to be the only independent predictor of rCDI.[19] Others have found lower antitoxin immunoglobulin levels at various times following the onset of iCDI to be predictive of a recurrence.[20, 21] A disadvantage of using these specialized tests as tools for clinical prediction is that they are not widely available in clinical practice. Even if these tests are available, their results are likely to return only after iCDI treatment has commenced. To make risk stratification more generalizable, we specifically focused on common data available in all clinical settings at the onset of iCDI.

We chose to restrict our risk stratification to factors present at the onset of iCDI for several reasons. First, earlier identification of patients at increased risk for rCDI may encourage clinicians to minimize subsequent exposures to non‐CDI antimicrobials and gastric acid suppressors. Second, newer anticlostridial therapies in development appear to target specifically CDI recurrence. The first anti‐CDI drug to be approved in 2 decades, fidaxomicin, has been shown to reduce the risk of a recurrence by nearly one‐half compared to vancomycin.[10, 11] Although in practice it is tempting to reserve this treatment for those patients who have multiple recurrences, there is no convincing evidence to date that the drug is similarly effective at reducing further recurrences in this population.[22, 23] Currently, the only population in which fidaxomicin treatment has been shown to reduce the risk of rCDI contains patients with at most 1 prior episode, whose first anti‐CDI exposure was to fidaxomicin.[10, 11] Thus, the intent of our model was to insure appropriate use of these new technologies from the perspective of both under‐ and overtreatment.

In general, most of the factors included in our model are neither novel nor surprising, including concurrent antibiotics and gastric acid suppression.[24, 25, 26, 27, 28, 29, 30] What is interesting about these exposures, however, is the fact that we measured them only at the onset of the iCDI episode. This implies that it is not merely the continuation of these medications after onset, but even exposure to them prior to the initial bout of CDI, that may promote a recurrence. This finding should give pause to the widespread practice of routinely prescribing gastric acid suppression to many hospitalized patients. It should also prompt a reexamination of antimicrobial choices for patients admitted for the treatment of infectious diseases in favor of those deemed at low risk for CDI whenever possible.

A relatively novel risk factor emerging from our model is the designation of the iCDI episode as CO‐HCFA.[30] A likely explanation for this relationship is that CO‐HCFA identifies a population of patients who are more ill, as evidenced by their prior hospitalization history. However, because recent hospitalizations themselves emerged as an independent predictor of rCDI in our model, CO‐HCFA designation clearly incorporates other factors important to this outcome.

Our data on illness severity are divergent from prior results. Previous work has found that increasing severity of illness is positively associated with the risk of a recurrence.[21, 31] In contrast, we found that the need for the ICU at the onset of iCDI appeared protective from rCDI. There are several explanations for this finding, the most likely being the competing mortality risk. Although we excluded from the study those patients who did not survive their iCDI hospitalizations, patients who received care in an ICU were more likely to die in the rCDI risk period than patients who did not receive care in an ICU (data not shown). Another potential explanation for this observation is that patients who develop iCDI while in the ICU may generally get more aggressive care than those contracting it on other wards, resulting in a lower risk for recurrence.

The recurrence rate in the current study is at the lower limit of what has been reported previously either in the meta‐analysis by Garey (13%50%) or in recent randomized controlled trials (25%).[10, 11, 25] This is likely due to our case identification pathway, and ascertainment bias is a potential limitation of our study. Patients with mild recurrent CDI diagnosed and treated as outpatients were not captured in our study unless their toxin assay was performed by the BJH laboratory (approximately 15% of specimens submitted to the BJH microbiology laboratory come from outpatients or affiliated outpatient or skilled nursing facilities). Similarly, recurrences diagnosed at other inpatient facilities were not captured in our study unless they were transferred to BJH for care. On the other hand, rCDI in randomized trials may be subject to a detection bias, because enrolled patients are prospectively monitored for and instructed to seek testing for recurrent diarrhea.

Our study also has limitations inherent to observational data such as confounding. We adjusted for all the available relevant potential confounders in the regression model. However, the possibility of residual confounding remains. Because our cohort was too small for a split‐cohort model validation, we employed a bootstrap method to cross‐validate our results. However, the model requires further validation in a prospective cohort in the future. The biggest limitation of our model, however, is its generalizability, because the data reflect patients and treatment patterns at an urban academic medical center, and may not mirror those of institutions with different characteristics or patients with iCDI diagnosed and managed completely in the outpatient setting.

In summary, we have developed a model to predict iCDI patients' risk of recurrence. The advantage of our model is the availability of all the factors at the onset of iCDI, when treatment decisions need to be made. Although far from perfect in its ability to discriminate those who will from those who will not develop a recurrence, it should serve as a beginning step in the direction of appropriately aggressive care that may result not only in diminishing the pool of this infection, but also in containing its spiraling costs. The cost‐benefit balance of these decisions needs to be examined explicitly, not only in terms of the financial cost of over‐ or undertreatment, but with respect to the implications of such overtreatment on development of resistance to newer anticlostridial agents.

Clostridium difficile infection (CDI) is a serious and costly condition whose volume in US hospitals has doubled over the last decade.[1, 2, 3] Along with this rise in incidence, its severity has also increased. Although in the United States there has been a doubling in age‐adjusted case fatality, in the same time period Canadian studies reported a high and increasing CDI‐associated case fatality in the setting of an outbreak of a novel epidemic hypervirulent strain BI/NAP1/027.[2, 4, 5, 6] The costs of CDI range widely ($2500 to $13,000 per hospitalization), with cumulative annual cost to the US healthcare system estimated at nearly $5 billion.[7, 8, 9]

One of the drivers of these clinical and economic outcomes is CDI recurrence (rCDI). In 2 recent randomized controlled trials, up to 25% of patients with an initial CDI (iCDI) episode developed rCDI.[10, 11] There are few data that quantify the impact of rCDI on quality of life and survival. However, patients often are readmitted to the hospital with rCDI, and physicians who treat patients with multiple episodes of rCDI can attest to the devastating toll it takes on the lives of the patients and their families (personal communications from numerous patients to E.R.D.).[12] Reducing the incidence of rCDI may significantly improve the course of this disease.

The advent of such new treatments as fidaxomicin aimed at rCDI is promising.[10, 11] However, evidence for its efficacy so far is limited to treatment‐naive iCDI patients, thus challenging clinicians to identify patients at high risk for rCDI at iCDI onset. To address this challenge, we set out to develop a bedside prediction model for rCDI based on the factors present and routinely available at the onset of iCDI.

METHODS

Study Design and Data Source

We conducted a retrospective single‐center cohort study to examine the factors present at the onset of iCDI that impact the incidence of rCDI among hospitalized patients. Patients were included in the study if they were adults (18 years) hospitalized at Barnes‐Jewish Hospital (BJH), St. Louis, Missouri, between January 1, 2003 and December 31, 2009, and who had a positive toxin assay for C difficile in the setting of unformed stools and no history of CDI in the previous 60 days (as defined by positive toxin assay). Patients were excluded if they either died during or were discharged to hospice from the iCDI hospitalization. Cases of iCDI were categorized according to published surveillance definitions as community onset‐healthcare facility associated (CO‐HCFA), healthcare facility onset, and community associated.[13] Notably, the CO‐HCFA category included surveillance definitions for both CO‐HCFA and indeterminate cases. We defined rCDI as a repeat positive toxin within 42 days following the end of iCDI treatment. This period of risk for rCDI was chosen because the current surveillance definition for rCDI is a new episode of CDI occurring within 8 weeks from the last episode of CDI, with the assumption the patient would receive 10 to 14 days of CDI treatment at the beginning of the 8‐week period.[14] Medical charts were reviewed for all readmissions during the recurrence risk period to identify patients diagnosed with rCDI by methods other than toxin assay. A study enrollment flow chart is shown in Figure 1.

Figure 1

Demographic and clinical data were derived from the BJH medical informatics databases and the BJH electronic medical records (see Supporting Appendix Table 1 in the online version of this article). Comorbidities were grouped using the Charlson‐Deyo categories.[15] All variables were limited to data that are consistent throughout a hospitalization (eg, race or age) or were present within 48 hours of iCDI (eg, medications).

RESULTS

Among the 4196 patients with iCDI enrolled in the study, 425 (10.1%) developed at least 1 recurrence within 42 days of the end of iCDI treatment (Table 1). Compared to patients without a recurrence, in univariate analysis those with an rCDI episode were older and had a greater comorbidity burden. In particular, diabetes mellitus (odds ratio 1.34; 95% confidence interval [CI], 1.08‐1.66) and cerebrovascular disease (odds ratio 1.47; 95% CI, 1.04‐2.08) were significantly more prevalent in the rCDI group. The index CDI episode for patients with rCDI was approximately twice as likely to fit the surveillance definition for CO‐HCFA than the index episode for those without a recurrence (odds ratio 2.24; 95% CI, 1.80‐2.79). Commensurately, patients with rCDI also had greater odds for experiencing multiple recent hospitalizations than those without rCDI. Neither type of CDI treatment (oral metronidazole vs oral vancomycin vs both), nor duration, was significantly associated with recurrence.

Seven factors present at the onset of iCDI were found to predict a recurrence in multivariable analysis (Table 2). Older age, CO‐HCFA status of iCDI, and 2 or more hospitalizations in the prior 60 days increased the risk of rCDI. Concomitant exposures to gastric acid suppressors, fluoroquinolone antibiotics, and high‐risk antibiotics were also significantly associated with a recurrence. Being in the intensive care unit (ICU) at the onset of iCDI was protective against rCDI in the multivariable model. This model had a C statistic of 0.642 and a Brier score of 0.089. After cross‐validation with 500 bootstrapping iterations, the model exhibited a moderately good fit (Figure 2). The prediction was particularly accurate in the lower risk ranges, with slight divergence in the risk strata over 20%. The validated model had a C statistic of 0.630 and Brier score of 0.089.

Figure 2

The sensitivity, specificity, and positive and negative predictive values of the model at various probability thresholds of rCDI are presented in Table 3. Thus, when the probability of rCDI was low, the model exhibited high sensitivity and low specificity. The situation was reversed as the probability of rCDI approached 30% (very low sensitivity and high specificity). The model's performance was optimal when the rCDI risk matched that in the current cohort, or 10.1%, with a sensitivity of 56% and specificity of 65%. However, when the rCDI risk dropped to 5%, the specificity dropped to below 30%. The sensitivity dropped to below 30% when rCDI risk rose to 15% (Table 3). Across the entire range of the probabilities tested, the negative predictive value of the model was persistently 90% or higher.

DISCUSSION

We have demonstrated that in a cohort of hospitalized patients with iCDI, 10% developed at least 1 episode of rCDI within 42 days of the end of iCDI treatment. The factors present at iCDI onset that predicted recurrence were age, CO‐HCFA CDI, prior hospitalization, high‐risk antibiotic and fluoroquinolone use, and gastric acid suppression. Although the model's performance was only moderate, its negative predictive value was 90% or higher across the entire range of rCDI probabilities tested. This means that the absence of this combination of risk factors in a patient with iCDI diminishes the probability of a rCDI episode to 10% or below, depending on the prior population risk for rCDI.

Prior investigators have developed prediction rules for rCDI. Hebert et al., using methodology similar to ours, constructed a model to predict the risk of rCDI among patients hospitalized with iCDI.[17] For example, the recurrence rate in their study was 23% compared to our 10%. This is likely due to the differing definitions of both iCDI and rCDI between the 2 studies. Although our definitions of hospital‐associated C difficile‐associated diarrhea (CDAD) conformed to the recommended surveillance definitions,[13] Hebert and colleagues used different definitions.[18] If this is so, the higher rate of rCDI in their study may have reflected these differences in surveillance definition, rather than the true prevalence of recurrent CDAD.

Several other studies have relied on either specialized laboratory tests alone or in combination with clinical factors. Stewart et al., in a small single‐center cohort study, reported the presence of the binary toxin to be the only independent predictor of rCDI.[19] Others have found lower antitoxin immunoglobulin levels at various times following the onset of iCDI to be predictive of a recurrence.[20, 21] A disadvantage of using these specialized tests as tools for clinical prediction is that they are not widely available in clinical practice. Even if these tests are available, their results are likely to return only after iCDI treatment has commenced. To make risk stratification more generalizable, we specifically focused on common data available in all clinical settings at the onset of iCDI.

We chose to restrict our risk stratification to factors present at the onset of iCDI for several reasons. First, earlier identification of patients at increased risk for rCDI may encourage clinicians to minimize subsequent exposures to non‐CDI antimicrobials and gastric acid suppressors. Second, newer anticlostridial therapies in development appear to target specifically CDI recurrence. The first anti‐CDI drug to be approved in 2 decades, fidaxomicin, has been shown to reduce the risk of a recurrence by nearly one‐half compared to vancomycin.[10, 11] Although in practice it is tempting to reserve this treatment for those patients who have multiple recurrences, there is no convincing evidence to date that the drug is similarly effective at reducing further recurrences in this population.[22, 23] Currently, the only population in which fidaxomicin treatment has been shown to reduce the risk of rCDI contains patients with at most 1 prior episode, whose first anti‐CDI exposure was to fidaxomicin.[10, 11] Thus, the intent of our model was to insure appropriate use of these new technologies from the perspective of both under‐ and overtreatment.

In general, most of the factors included in our model are neither novel nor surprising, including concurrent antibiotics and gastric acid suppression.[24, 25, 26, 27, 28, 29, 30] What is interesting about these exposures, however, is the fact that we measured them only at the onset of the iCDI episode. This implies that it is not merely the continuation of these medications after onset, but even exposure to them prior to the initial bout of CDI, that may promote a recurrence. This finding should give pause to the widespread practice of routinely prescribing gastric acid suppression to many hospitalized patients. It should also prompt a reexamination of antimicrobial choices for patients admitted for the treatment of infectious diseases in favor of those deemed at low risk for CDI whenever possible.

A relatively novel risk factor emerging from our model is the designation of the iCDI episode as CO‐HCFA.[30] A likely explanation for this relationship is that CO‐HCFA identifies a population of patients who are more ill, as evidenced by their prior hospitalization history. However, because recent hospitalizations themselves emerged as an independent predictor of rCDI in our model, CO‐HCFA designation clearly incorporates other factors important to this outcome.

Our data on illness severity are divergent from prior results. Previous work has found that increasing severity of illness is positively associated with the risk of a recurrence.[21, 31] In contrast, we found that the need for the ICU at the onset of iCDI appeared protective from rCDI. There are several explanations for this finding, the most likely being the competing mortality risk. Although we excluded from the study those patients who did not survive their iCDI hospitalizations, patients who received care in an ICU were more likely to die in the rCDI risk period than patients who did not receive care in an ICU (data not shown). Another potential explanation for this observation is that patients who develop iCDI while in the ICU may generally get more aggressive care than those contracting it on other wards, resulting in a lower risk for recurrence.

The recurrence rate in the current study is at the lower limit of what has been reported previously either in the meta‐analysis by Garey (13%50%) or in recent randomized controlled trials (25%).[10, 11, 25] This is likely due to our case identification pathway, and ascertainment bias is a potential limitation of our study. Patients with mild recurrent CDI diagnosed and treated as outpatients were not captured in our study unless their toxin assay was performed by the BJH laboratory (approximately 15% of specimens submitted to the BJH microbiology laboratory come from outpatients or affiliated outpatient or skilled nursing facilities). Similarly, recurrences diagnosed at other inpatient facilities were not captured in our study unless they were transferred to BJH for care. On the other hand, rCDI in randomized trials may be subject to a detection bias, because enrolled patients are prospectively monitored for and instructed to seek testing for recurrent diarrhea.

Our study also has limitations inherent to observational data such as confounding. We adjusted for all the available relevant potential confounders in the regression model. However, the possibility of residual confounding remains. Because our cohort was too small for a split‐cohort model validation, we employed a bootstrap method to cross‐validate our results. However, the model requires further validation in a prospective cohort in the future. The biggest limitation of our model, however, is its generalizability, because the data reflect patients and treatment patterns at an urban academic medical center, and may not mirror those of institutions with different characteristics or patients with iCDI diagnosed and managed completely in the outpatient setting.

In summary, we have developed a model to predict iCDI patients' risk of recurrence. The advantage of our model is the availability of all the factors at the onset of iCDI, when treatment decisions need to be made. Although far from perfect in its ability to discriminate those who will from those who will not develop a recurrence, it should serve as a beginning step in the direction of appropriately aggressive care that may result not only in diminishing the pool of this infection, but also in containing its spiraling costs. The cost‐benefit balance of these decisions needs to be examined explicitly, not only in terms of the financial cost of over‐ or undertreatment, but with respect to the implications of such overtreatment on development of resistance to newer anticlostridial agents.

Clostridium difficile infection (CDI) is a serious and costly condition whose volume in US hospitals has doubled over the last decade.[1, 2, 3] Along with this rise in incidence, its severity has also increased. Although in the United States there has been a doubling in age‐adjusted case fatality, in the same time period Canadian studies reported a high and increasing CDI‐associated case fatality in the setting of an outbreak of a novel epidemic hypervirulent strain BI/NAP1/027.[2, 4, 5, 6] The costs of CDI range widely ($2500 to $13,000 per hospitalization), with cumulative annual cost to the US healthcare system estimated at nearly $5 billion.[7, 8, 9]

One of the drivers of these clinical and economic outcomes is CDI recurrence (rCDI). In 2 recent randomized controlled trials, up to 25% of patients with an initial CDI (iCDI) episode developed rCDI.[10, 11] There are few data that quantify the impact of rCDI on quality of life and survival. However, patients often are readmitted to the hospital with rCDI, and physicians who treat patients with multiple episodes of rCDI can attest to the devastating toll it takes on the lives of the patients and their families (personal communications from numerous patients to E.R.D.).[12] Reducing the incidence of rCDI may significantly improve the course of this disease.

The advent of such new treatments as fidaxomicin aimed at rCDI is promising.[10, 11] However, evidence for its efficacy so far is limited to treatment‐naive iCDI patients, thus challenging clinicians to identify patients at high risk for rCDI at iCDI onset. To address this challenge, we set out to develop a bedside prediction model for rCDI based on the factors present and routinely available at the onset of iCDI.

METHODS

Study Design and Data Source

We conducted a retrospective single‐center cohort study to examine the factors present at the onset of iCDI that impact the incidence of rCDI among hospitalized patients. Patients were included in the study if they were adults (18 years) hospitalized at Barnes‐Jewish Hospital (BJH), St. Louis, Missouri, between January 1, 2003 and December 31, 2009, and who had a positive toxin assay for C difficile in the setting of unformed stools and no history of CDI in the previous 60 days (as defined by positive toxin assay). Patients were excluded if they either died during or were discharged to hospice from the iCDI hospitalization. Cases of iCDI were categorized according to published surveillance definitions as community onset‐healthcare facility associated (CO‐HCFA), healthcare facility onset, and community associated.[13] Notably, the CO‐HCFA category included surveillance definitions for both CO‐HCFA and indeterminate cases. We defined rCDI as a repeat positive toxin within 42 days following the end of iCDI treatment. This period of risk for rCDI was chosen because the current surveillance definition for rCDI is a new episode of CDI occurring within 8 weeks from the last episode of CDI, with the assumption the patient would receive 10 to 14 days of CDI treatment at the beginning of the 8‐week period.[14] Medical charts were reviewed for all readmissions during the recurrence risk period to identify patients diagnosed with rCDI by methods other than toxin assay. A study enrollment flow chart is shown in Figure 1.

Figure 1

Demographic and clinical data were derived from the BJH medical informatics databases and the BJH electronic medical records (see Supporting Appendix Table 1 in the online version of this article). Comorbidities were grouped using the Charlson‐Deyo categories.[15] All variables were limited to data that are consistent throughout a hospitalization (eg, race or age) or were present within 48 hours of iCDI (eg, medications).

RESULTS

Among the 4196 patients with iCDI enrolled in the study, 425 (10.1%) developed at least 1 recurrence within 42 days of the end of iCDI treatment (Table 1). Compared to patients without a recurrence, in univariate analysis those with an rCDI episode were older and had a greater comorbidity burden. In particular, diabetes mellitus (odds ratio 1.34; 95% confidence interval [CI], 1.08‐1.66) and cerebrovascular disease (odds ratio 1.47; 95% CI, 1.04‐2.08) were significantly more prevalent in the rCDI group. The index CDI episode for patients with rCDI was approximately twice as likely to fit the surveillance definition for CO‐HCFA than the index episode for those without a recurrence (odds ratio 2.24; 95% CI, 1.80‐2.79). Commensurately, patients with rCDI also had greater odds for experiencing multiple recent hospitalizations than those without rCDI. Neither type of CDI treatment (oral metronidazole vs oral vancomycin vs both), nor duration, was significantly associated with recurrence.

Seven factors present at the onset of iCDI were found to predict a recurrence in multivariable analysis (Table 2). Older age, CO‐HCFA status of iCDI, and 2 or more hospitalizations in the prior 60 days increased the risk of rCDI. Concomitant exposures to gastric acid suppressors, fluoroquinolone antibiotics, and high‐risk antibiotics were also significantly associated with a recurrence. Being in the intensive care unit (ICU) at the onset of iCDI was protective against rCDI in the multivariable model. This model had a C statistic of 0.642 and a Brier score of 0.089. After cross‐validation with 500 bootstrapping iterations, the model exhibited a moderately good fit (Figure 2). The prediction was particularly accurate in the lower risk ranges, with slight divergence in the risk strata over 20%. The validated model had a C statistic of 0.630 and Brier score of 0.089.

Figure 2

The sensitivity, specificity, and positive and negative predictive values of the model at various probability thresholds of rCDI are presented in Table 3. Thus, when the probability of rCDI was low, the model exhibited high sensitivity and low specificity. The situation was reversed as the probability of rCDI approached 30% (very low sensitivity and high specificity). The model's performance was optimal when the rCDI risk matched that in the current cohort, or 10.1%, with a sensitivity of 56% and specificity of 65%. However, when the rCDI risk dropped to 5%, the specificity dropped to below 30%. The sensitivity dropped to below 30% when rCDI risk rose to 15% (Table 3). Across the entire range of the probabilities tested, the negative predictive value of the model was persistently 90% or higher.

DISCUSSION

We have demonstrated that in a cohort of hospitalized patients with iCDI, 10% developed at least 1 episode of rCDI within 42 days of the end of iCDI treatment. The factors present at iCDI onset that predicted recurrence were age, CO‐HCFA CDI, prior hospitalization, high‐risk antibiotic and fluoroquinolone use, and gastric acid suppression. Although the model's performance was only moderate, its negative predictive value was 90% or higher across the entire range of rCDI probabilities tested. This means that the absence of this combination of risk factors in a patient with iCDI diminishes the probability of a rCDI episode to 10% or below, depending on the prior population risk for rCDI.

Prior investigators have developed prediction rules for rCDI. Hebert et al., using methodology similar to ours, constructed a model to predict the risk of rCDI among patients hospitalized with iCDI.[17] For example, the recurrence rate in their study was 23% compared to our 10%. This is likely due to the differing definitions of both iCDI and rCDI between the 2 studies. Although our definitions of hospital‐associated C difficile‐associated diarrhea (CDAD) conformed to the recommended surveillance definitions,[13] Hebert and colleagues used different definitions.[18] If this is so, the higher rate of rCDI in their study may have reflected these differences in surveillance definition, rather than the true prevalence of recurrent CDAD.

Several other studies have relied on either specialized laboratory tests alone or in combination with clinical factors. Stewart et al., in a small single‐center cohort study, reported the presence of the binary toxin to be the only independent predictor of rCDI.[19] Others have found lower antitoxin immunoglobulin levels at various times following the onset of iCDI to be predictive of a recurrence.[20, 21] A disadvantage of using these specialized tests as tools for clinical prediction is that they are not widely available in clinical practice. Even if these tests are available, their results are likely to return only after iCDI treatment has commenced. To make risk stratification more generalizable, we specifically focused on common data available in all clinical settings at the onset of iCDI.

We chose to restrict our risk stratification to factors present at the onset of iCDI for several reasons. First, earlier identification of patients at increased risk for rCDI may encourage clinicians to minimize subsequent exposures to non‐CDI antimicrobials and gastric acid suppressors. Second, newer anticlostridial therapies in development appear to target specifically CDI recurrence. The first anti‐CDI drug to be approved in 2 decades, fidaxomicin, has been shown to reduce the risk of a recurrence by nearly one‐half compared to vancomycin.[10, 11] Although in practice it is tempting to reserve this treatment for those patients who have multiple recurrences, there is no convincing evidence to date that the drug is similarly effective at reducing further recurrences in this population.[22, 23] Currently, the only population in which fidaxomicin treatment has been shown to reduce the risk of rCDI contains patients with at most 1 prior episode, whose first anti‐CDI exposure was to fidaxomicin.[10, 11] Thus, the intent of our model was to insure appropriate use of these new technologies from the perspective of both under‐ and overtreatment.

In general, most of the factors included in our model are neither novel nor surprising, including concurrent antibiotics and gastric acid suppression.[24, 25, 26, 27, 28, 29, 30] What is interesting about these exposures, however, is the fact that we measured them only at the onset of the iCDI episode. This implies that it is not merely the continuation of these medications after onset, but even exposure to them prior to the initial bout of CDI, that may promote a recurrence. This finding should give pause to the widespread practice of routinely prescribing gastric acid suppression to many hospitalized patients. It should also prompt a reexamination of antimicrobial choices for patients admitted for the treatment of infectious diseases in favor of those deemed at low risk for CDI whenever possible.

A relatively novel risk factor emerging from our model is the designation of the iCDI episode as CO‐HCFA.[30] A likely explanation for this relationship is that CO‐HCFA identifies a population of patients who are more ill, as evidenced by their prior hospitalization history. However, because recent hospitalizations themselves emerged as an independent predictor of rCDI in our model, CO‐HCFA designation clearly incorporates other factors important to this outcome.

Our data on illness severity are divergent from prior results. Previous work has found that increasing severity of illness is positively associated with the risk of a recurrence.[21, 31] In contrast, we found that the need for the ICU at the onset of iCDI appeared protective from rCDI. There are several explanations for this finding, the most likely being the competing mortality risk. Although we excluded from the study those patients who did not survive their iCDI hospitalizations, patients who received care in an ICU were more likely to die in the rCDI risk period than patients who did not receive care in an ICU (data not shown). Another potential explanation for this observation is that patients who develop iCDI while in the ICU may generally get more aggressive care than those contracting it on other wards, resulting in a lower risk for recurrence.

The recurrence rate in the current study is at the lower limit of what has been reported previously either in the meta‐analysis by Garey (13%50%) or in recent randomized controlled trials (25%).[10, 11, 25] This is likely due to our case identification pathway, and ascertainment bias is a potential limitation of our study. Patients with mild recurrent CDI diagnosed and treated as outpatients were not captured in our study unless their toxin assay was performed by the BJH laboratory (approximately 15% of specimens submitted to the BJH microbiology laboratory come from outpatients or affiliated outpatient or skilled nursing facilities). Similarly, recurrences diagnosed at other inpatient facilities were not captured in our study unless they were transferred to BJH for care. On the other hand, rCDI in randomized trials may be subject to a detection bias, because enrolled patients are prospectively monitored for and instructed to seek testing for recurrent diarrhea.

Our study also has limitations inherent to observational data such as confounding. We adjusted for all the available relevant potential confounders in the regression model. However, the possibility of residual confounding remains. Because our cohort was too small for a split‐cohort model validation, we employed a bootstrap method to cross‐validate our results. However, the model requires further validation in a prospective cohort in the future. The biggest limitation of our model, however, is its generalizability, because the data reflect patients and treatment patterns at an urban academic medical center, and may not mirror those of institutions with different characteristics or patients with iCDI diagnosed and managed completely in the outpatient setting.

In summary, we have developed a model to predict iCDI patients' risk of recurrence. The advantage of our model is the availability of all the factors at the onset of iCDI, when treatment decisions need to be made. Although far from perfect in its ability to discriminate those who will from those who will not develop a recurrence, it should serve as a beginning step in the direction of appropriately aggressive care that may result not only in diminishing the pool of this infection, but also in containing its spiraling costs. The cost‐benefit balance of these decisions needs to be examined explicitly, not only in terms of the financial cost of over‐ or undertreatment, but with respect to the implications of such overtreatment on development of resistance to newer anticlostridial agents.