Hospital readmissions: prediction and inference

Hospital readmission rates vary widely across regions and hospitals, suggesting that improvements are possible. To provide an incentive to improve quality of care, some jurisdictions have introduced legislation that financially penalizes hospitals with high readmission rates, but the models used to implement the legislation crudely adjusts for patient-mix. Additionally, hospitals have developed predictive models of readmission risk to better target enhanced transitional care. In this work, I examined how large healthcare administrative databases can help build better inferential and predictive models of hospital readmissions. To target interventions at those patients with the highest readmission risk, hospitals can develop predictive models of readmission based on their own data (local models), they can pool their data with other hospitals (global models) or they can use sophisticated model combination techniques which avoid directly sharing patient data (combined models). In the first manuscript, I compared the accuracy of global, combined, and local models in predicting 30-day readmission risk, and found that the predictive accuracy of models developed with the three approaches were similar, suggesting that hospitals can use their own data to accurately predict hospital readmissions. Although predictive models of hospital readmissions can be useful to guide resources to individual high-risk cases, inferential models can potentially lead to population-level interventions. In the second manuscript, I studied how the day-of-week of discharge affects readmission, and used both empiric (survival model) and analytic (Markov model) approaches to study how this effect is confounded by the probability of admission on the weekend. I found that not only are Friday discharges more likely to be readmitted than Wednesday discharges, but also that the low probability of weekend admissions attenuates this effect if uncontrolled. Our results suggest that interventions that reduce the effect of Friday discharge on readmissions, such as increased weekend staffing, are likely to be more cost-effective than previous work has indicated.In the third manuscript, I compare two techniques to measure the effect of twenty Montreal hospitals on readmissions: a standard regression approach that controls for the major, well-known confounders, and targeted maximum likelihood estimation (TMLE) where I could control for pre-admission diagnoses, procedures, and drug prescriptions using a machine learning technique (random forest). The standard model suggested that there was little difference between the hospitals, but the TMLE model showed that the confounders, particularly drug prescriptions, strongly confounded readmission risk, and revealed a wide variation in readmission risk between the hospitals. My work suggests that: 1) predictive models of readmission are unlikely to be greatly improved by pooling hospital data or by using complex combination techniques, 2) inference on the causes of readmissions, particularly the day-of-week, can be confounded by the admission process, and 3) by using TMLE, the predictive power of machine learning techniques can be used to improve inference by reducing bias in our estimates of the effect of hospital care on readmissions.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.