Prospective evaluation of social risks, physical function, and cognitive function in prediction of non-elective rehospitalization and post-discharge mortality

Increasing evidence suggests social factors like unstable housing [1]; food insecurity [2‐7]; transportation difficulties [8]; chronic stress [9]; and ability to get help with activities of daily living [10‐13] influence various health and healthcare utilization outcomes, including risk of hospitalization and frequent use of the emergency department (ED) [14]. Similarly, problems with physical and cognitive function may contribute to patients’ rehospitalization risk [15, 16]. Understanding a patient’s risk for readmission may help healthcare providers develop appropriate tailored treatment and post-discharge care plans to reduce potentially avoidable readmissions and mortality. Teasing out the independent contributions of these factors to health outcomes is complicated since they may coexist and correlate with illness or disability burden.

In 2015, we described the development of a real-time score – referred to as the Transitions Support Level (TSL) score – to predict rehospitalization and post-discharge mortality risk [17]. Subsequently, Kaiser Permanente Northern California (KPNC) hospitals instantiated this model in the electronic health record (EHR). The TSL employs longitudinal comorbidity, acute physiology, discharge care directives, and recent utilization [17] and generates a predicted percent risk for a composite outcome (non-elective hospitalization and/or death within 30 days following hospital discharge) ranging from 0 to 100%. Risk estimates based on the TSL were integrated with discharge and follow-up workflows, which has resulted in decreased rehospitalizations [18]. In that same year Kaiser Permanente’s (KP)’s Care Management Institute developed the Your Current Life Situation (YCLS) questionnaire and item bank to screen members for social factors that could affect health and healthcare access [19]. YCLS items have been used in a variety of pilot projects in different programs and settings but are not part of routine outpatient or inpatient assessment. During this time period, KPNC also started to evaluate the use of brief patient-reported measures of functional status.

To test the hypothesis that patient-reported information could enhance the TLS’s prediction of post-hospital outcomes over EHR data alone, we employed a prospective cohort study design. The aim of our study was to test the hypothesis that adding patient-reported data about social factors and physical and cognitive function to the TLS data would significantly enhance the ability to predict 30-day rehospitalization and mortality. A secondary aim was to determine what specific patient-reported information might be recommended for routine ascertainment and entry into the EHR.

During the study recruitment period, 36,086 patients were hospitalized at the three participating sites. Of these patients, 4,811 met all initial screening and eligibility criteria as determined by the research staff. Of these, 2,263 patients were approached by research staff. Of these patients, 1,551 agreed to be interviewed and 712 declined to participate. The remaining 2,548 patients were excluded or not selected (i.e., not approached by research staff). See Fig. 1 for additional context.

During the recruitment period there were 36,086 adult patients hospitalized at the three Kaiser Permanente Northern California participating Medical Centers with an inpatient or observation designation. 2,548 patients not approached for various reasons were categorized as either “excluded” (N = 1,171) or “not selected” (N = 1,377). Excluded patients met most in-hospital eligibility criteria, but were not approached for another reason (i.e., nurse disapproved approaching patient, patient was already discharged, time constraints, or other as noted by research staff – such as “patient sleeping” or “family member visiting”). Patients “not selected” included those determined to be ineligible when research staff reviewed the patient’s electronic medical records on the day of recruitment. Not selected patients included: patients with discharge location unknown; patient continuing care at another facility after discharge; patient began their hospitalization at another non-KP facility and was transferred in; non-KP insurance coverage; comfort care only order in effect, could not confirm eligibility; or could not speak English. There were 1,551 patients interviewed and enrolled in the study. Four patients died prior to discharge and were removed from the analysis.

Table 1 shows the characteristics of patients during their selected hospitalization who were interviewed, excluded or not selected, and all other hospitalized patients (additional comparisons, Appendix 4).

The interviewed and enrolled cohort had more African American patients but fewer Hispanics. When compared to all other hospitalizations, interviewed patients were sicker, with higher CCI, COPS2, and LAPS2 scores. However, the interviewed cohort was healthier, with lower CCI, COPS2, and LAPS2 scores, than those who were excluded or not selected. Only 0.5% of the interviewed cohort was admitted for observation, and LOS was longer, but the distribution of discharge diagnoses was similar to that of the rest of the cohort. Four (0.3%) patients died in the hospital after their interview. In the remaining 1,547 interviewed patients, post-discharge mortality was somewhat lower than in the remaining patients. Compared to the rest of the hospitalized cohort, excluded patients’ TSL scores were higher (mean, 16.2, 11.4, p < 0.0001). Interviewed patients’ TSL scores were somewhat higher than the rest of the hospitalized cohort (12.8 vs. 11.4, p < 0.0001). Table 2 shows that, due to logistic difficulties, we were not successful in oversampling patients with TSL scores ≥ 15%. However, the table shows that the TSL score distribution of the interviewed patients is similar to that of all adult hospitalizations prior to the study and that weighting the sample was not necessary.

Of the 1,547 interviewed patients who left the hospital alive, 212 (13.7%) experienced the adverse composite outcome. Table 3 shows that interviewed patients’ cognitive function (mean, 32.4) was outside the normalized PROMIS range, but their physical function (mean 53.6) was within the normalized PROMIS range.

Table 3 shows the two most common social risk factors as presence of a health problem or disability (51.3%) and not living with a partner (47.1%); the prevalence of the remaining social risk factors ranged from 8.4%-15.8%. Table 3 also compares the 212 patients who experienced the composite outcome with the 1335 patients who did not. Patients with the composite outcome had significantly higher CCI, COPS2, LAPS2, and TSL scores (all comparisons, p < 0.0001), and their cognitive function (30.8 vs. 31.8, p = 0.027) and physical function (52.3 vs. 53.9, p = 0.01) scores were lower than those who did not experience the outcome. Patients with the composite outcome were significantly more likely to have transportation difficulties (22.8% vs. 14.8%, p = 0.008) and to have a health problem or disability that affected their activities of daily living (66.0% vs. 49.0%, p < 0.001). Among patients with a health or disability issue, those with the composite outcome were more likely to report having help readily available.

Table 4 shows the performance of different multivariable models for predicting the composite outcome (Appendix 5 shows model coefficients). We found that incorporating the social factors and PROMIS variables did not improve model discrimination or explanatory power over the health and demographic data routinely captured in the EHR. Importantly, a random forest model that included all available predictors did not result in improved performance as compared to the logistic model with TSL. Analyses restricted to patients with mid-range (15.0–44.9%) TSL scores showed similar results.

Correlations across continuous variables (age, LAPS2, COPS2, TSL, and PROMIS) and cross tabulations between these variables and individual social factor variables are provided in Appendix 3. Correlations between the 2 PROMIS scales and the other continuous variables were statistically significant but not necessarily clinically significant, with the highest correlation involving the PROMIS scales found for cognitive function and age (-0.15, p < 0.001). Age was positively correlated with severity of illness (LAPS2), comorbidity burden (COPS2), and TSL score. Cross tabulations between the continuous variables and individual social factors variables did not show a consistent pattern except with respect to the PROMIS physical function scale, where all patients with unfavorable social factors had significantly lower physical function (all 6 comparisons with p < 0.001).

Our prospective study with hospitalized patients found that adding patient-reported information about social factors, physical function, and cognitive status did not improve the statistical performance of a predictive model of rehospitalization or death within 30 days post-discharge compared to a model based on EHR clinical data alone. However, we did find that presence of unfavorable patient-reported social risks, lower physical function, and lower cognitive function were independently associated with a higher rate of the composite outcome.

Although we cannot compare our findings to other studies directly, our results are consistent with the findings of Bhavsar et al., who studied 8 outcomes (ED use, hospitalization, outpatient visit use, occurrence of myocardial infarction, stroke, asthma, accidents, and influenza). Their team found that adding neighborhood socioeconomic status variables did not improve upon the statistical performance of predictive models using routinely available clinical data from an EHR [31]. Our findings are also consistent with our own recent analyses, where we found that adding neighborhood socioeconomic status variables did not improve prediction models for mortality and 6 health care utilization variables [32].

It is important to highlight key limitations of our study. First, our cohort was restricted to an insured population in a health system with a relatively high degree of vertical outpatient-inpatient service integration. This integration is manifest in the availability of services aimed at preventing hospitalization, such as chronic condition management programs, a central call center, electronic patient portals, and automated risk detection, which have resulted in a steady decrease in rehospitalization and post-discharge mortality [24]. Second, due to resource limitations, we could not interview patients on weekends, nor could we interview non-English speakers, which latter effect is most evident in the reduced proportion of Hispanic patients in our cohort. Because of state of California regulatory requirements, we could not interview very low- income Kaiser Foundation Health Plan members covered through Medi-Cal (California’s Medicaid program). Thus, it is possible that, because of our setting and selection process, our study sample may not be generalizable to other settings. Third, since social factors and measures of physical and cognitive function were all patient-reported, it is possible that people’s responses did not accurately reflect their true level of vulnerability. These aforementioned exclusions and limitations may have resulted in an overall lower risk hospitalized cohort, thereby reducing our ability for our measures to impact the score. Fourth, it is possible that confounding or redundancy factors were occurring. For example, patients reporting health or disability problems who also had readily available help from others were more likely to experience the adverse composite outcome. This could be because people with health or disability problems that limit their ability to take care of their daily needs are more likely to have help availability than those not so limited. Lastly, our sample might consist of very similar patients for several reasons making it more difficult for a model to discriminate patients by their future outcome. The sample comes from a capitated staff model managed care insured population where patients at the extremes of the social and functional scales might not be represented. We also sampled only low-mid risk patients, and the sample of interviewed patients appeared to be healthier and lower risk than those excluded from the study.

It is important to consider, however, that the most important factor affecting our results is that the outcome time frame (30 days after discharge) is very narrow. Most comorbid conditions develop over many years, and it seems likely that by the time a patient has been hospitalized, the effect of social factors and lower physical and cognitive function has already manifested as a set of comorbid conditions. Not surprisingly then, predictive models that include quantification of comorbidities and acute physiology are unlikely to be affected by including factors that may have existed for years prior to the hospitalization and predisposing patients to develop conditions resulting in hospitalization.

The fact that the patient-reported social factors and PROMIS cognitive and physical function status variables we employed failed to enhance prediction for a 30-day outcome does not mean that ascertaining this information is not valuable. Having this information available in the EHR at time of discharge planning could be of significant value in designing tailored interventions and in assessing which patients might benefit most from specific interventions [33]. For example, patients with social risks or current social needs could be triaged to KPNC Continuum of Care programs such as Hospital-to-Home Transitions and Complex Needs Programs for follow-up in the outpatient setting. Additionally, it is likely that identifying and addressing social risks such as access to healthy food, transportation, a stable and supportive living situation, instrumental and emotional social support, and health literacy will have a positive impact on the recovery process, including reduction in medical complications and potentially preventable ED and outpatient visits. We have shown that, when information on benefit from an intervention is available, this information can be used to develop more accurate rehospitalization predictive models [34]. These variables could enhance the predictive ability of other outcomes in different populations, so it is important to continue to evaluate and quantify the value of collecting these data in studies such as this one.

It is important to recognize another key strength of this study. We were able to successfully collect social factors, physical function, and cognitive status patient-reported information during an inpatient stay using standardized questionnaires administered via tablet. It is feasible to collect this data and knowing this information about a patient at discharge may allow for more complex and tailored interventions aimed at improving outcomes and reducing readmissions [35].