Predicting Hospital Readmission among Patients with Sepsis using Clinical and Wearable Data

I. INTRODUCTION

Sepsis is defined as a dysregulated host response to infection and remains a major public health burden [1]. Worldwide, approximately 50 million people develop sepsis each year and it is estimated that sepsis is responsible for 1 out of every 5 global deaths [2]. In the United States, it is estimated that over 1.7 million adults in America develop sepsis every year, resulting in over 350,000 deaths [3].

While there have been significant efforts to decrease mortality of sepsis, recent data demonstrate that sepsis survivors are at high risk of various complications, including prolonged immunosuppression, physical deconditioning and neurocognitive impairments [4]. This may explain why patients with sepsis have significantly higher rates of unplanned readmissions than patients with heart failure, myocardial infarction, pneumonia and chronic obstructive lung disease (COPD) [5], [6], [7]. Additionally, recent data suggests sepsis readmissions are responsible for over $3.5 billion dollars a year in healthcare expenditures in the United States alone [6]. Importantly, there have been only a handful of interventions that have shown promise to decrease unplanned sepsis readmissions [8]. The importance of care after discharge cannot be understated. A 10-year Taiwanese study of over 15,000 patients found a 5.6% mortality risk reduction in those who received rehabilitation within 3 months of discharge [9]. A German study found decreased mortality among patients who suffered from a severe infection or more and were sent to post-discharge rehabilitation [10].

Wearable devices such as Fitbit allow for tracking vital signs and physical activity levels have proliferated continuously in the past decades, but the involvement of wearable data within clinical applications remains scarce [11].

Nemati et al. developed a model to detect atrial fibrillation using photoplethysmogram (PPG) recorded from simband smartwatches with an area under the receiver operating curve (AUROC) of 0.99 [12]. W. J.Kane et al. used step count recorded by wearable devices and showed that 10% increase in step count 30 days after colorectal surgery decreased the risk of 30 days readmission by 40% [13]. Such devices may help identify patients at high risk for physiologic deterioration and potential readmission to the hospital. In this study, we hypothesized that wearable data are predictive factors for assessing the risk of rehospitalization among patients with sepsis.

II. METHODS

III. RESULTS & DISCUSSION

We identified 9,910 index admissions with sepsis, among which 2815 were readmitted in 90 days (29%). Table I summarizes characteristics of patients who were readmitted and not readmitted within 90 days of hospital discharge.

We identified 103 patients in the current release of the AllofUS dataset (2021Q3R6) with Fitbit recordings. Of these, thirty-nine of patients with sepsis had Fitbit data available during their hospitalization (pre-discharge records) or after hospital discharge (post-discharge records), 7 of whom were readmitted in 90 days. Figure 1 shows the association of sedentary hours per day and readmission risk. We found that patients who were more sedentary were more likely to be readmitted by 90-days post-discharge. Table II provides each model’s input and their performances.

• Download figure
• Open in new tab

Fig. 1.

An illustrative example of assessing feature distribution among patients with sepsis re-admitted vs non readmitted. Upper panels show the histogram of the sedentary hours of every day, recorded every minute, for readmitted vs non readmitted patients with sepsis. The lower panel shows the CDF for both population cohorts and the kolmogorov–Smirnov test (KS test) reveals that the distribution of activity hours for re-admitted vs non readmitted patients are dissimilar.

We found that Model 6, which included clinical data as well as pre- and post-discharge Fitbit data had the best ability to predict a readmission, including a PPV of 60% at 80% sensitivity. In comparison, Model 1 which only included pre-discharge clinical features had a PPV of 26% at 80% sensitivity.

Figure 2 illustrates top 5 predictor factors for the risk of readmission at day 1 and day 10 post-discharge. We found that at day 1 pre-discharge features such as Sedentary rate, O₂Sat, blood Glucose level, CRP (a biomarker of inflammation) and heart rate were significant predictors of readmission. By day 10, factors such as post-discharge Sedentary rate showed significant predictive value.

• Download figure
• Open in new tab

Fig. 2.

Shapley values of the top 5 predictive features in readmission risk. (A): Post discharge day 1 (B): Post discharge day 10. CRP = C-reactive protein, Sedentary= Sedentary minutes per day (Activity level).

There are few validated models that accurately predict readmission following an index hospitalization for sepsis. Commonly used tools, such as the LACE+ index and HOSPITAL score were not developed specifically for sepsis patients and perform poorly at predicting readmissions in this population [19]. A previous EHR-based model developed by our team [19], using only pre-discharge clinical features, demonstrated improved predictive power over LACE+, but the model AUCroc was around 0.71. An accurate model for assessing the risk of readmission may help healthcare providers allocate resources to higher risk patients and prevent unplanned readmissions.

While in this work we focused on readmission prediction, equally important is the problem of assessing discharge readiness, or more broadly figuring out if de-escalation of care-level is warranted. A comparison of Model II and Model IV indicate that wearable data are likely to contribute to de-escalation of care and discharge planning. Moreover, effective implementation of preventive strategies for reduction of unplanned readmission often requires involving a care-coordinator three or more days in advance of patient discharge. As noted by Kash1 and colleagues [20], some of the most effective interventions to reduce unplanned readmissions include better coordination among clinical teams and/or community providers, post-discharge home visits, telephone follow-up calls, patient/family education, and discharge planning. As such, continuous prediction of discharge readiness and prediction of unplanned readmission during hospitalization is likely to help with timely care-coordination in order to avoid unplanned readmission.

Fitbit or other wearable data can also provide monitoring for sedentary behaviors in the after-discharge period which is known to also be associated with increased mortality. On a national level, the Centers for Medicare and Medical Services (CMS) uses the rehospitalization rate to adjust the hospitals’ reimbursement in favor of hospitals with a lower readmission rate [21].

Additionally, sepsis readmissions carry a major economic burden and effective strategies to curtail this are lacking. Although our data are small, our preliminary results suggest there may be benefit of incorporating Fitbit data (or other wearable devices) into risk prediction models to decrease the economic burden of sepsis readmissions. This work developed different models for predicting the risk of read-mission for patients with sepsis based on their clinical and wearable data before and after discharge. Our major finding is that Model 6 (which uses clinical features, including pre-discharge and post-discharge wearable data) significantly outperformed the other models in all performance metrics. Importantly, we found that the model trained with post-discharge wearable data features had higher specificity and PPV than models that trained on other input features. We acknowledge several limitations to our work. First, data were retrospective and we are unable to demonstrate causation. Next, we only had a small number of patients with Fitbit data which limits our findings.

In summary, we found that models using wearable sensor data may have superior performance in predicting 90-day readmissions following an index sepsis hospitalization than those that do not utilize this data. Future, prospective studies should be conducted to evaluate the performance of these models.