Study tests ML for hospital readmission risk

A study published in PLOS ONE on June 5, 2026 examined how well machine-learning models can predict 30-day hospital readmission using electronic health record data. The work drew on Epic electronic health record data from a single academic medical center, encompassing adult inpatient admissions between April 2018 and February 2024. Within this cohort, the observed 30-day readmission rate was 9.9%, providing the baseline against which model performance was evaluated.

The investigators focused on routinely collected clinical information available in the Epic system. By restricting the analysis to one institution, they were able to apply a consistent data structure and coding framework across all admissions in the study period. This design allowed direct comparison of different predictive approaches using the same underlying dataset.

The research team compared several machine-learning algorithms, including XGBoost and deep neural networks, with traditional logistic regression for predicting readmission. Overall model discrimination was modest. Across evaluated models, the area under the receiver operating characteristic curve (AUROC) ranged from 0.62 to 0.64, indicating limited ability to distinguish between patients who would be readmitted within 30 days and those who would not.

Averaged performance metrics across models showed a sensitivity of 0.64, with a range from 0.32 to 0.92. The average positive predictive value was 0.16, with a range of 0.12 to 0.22, meaning that only a small fraction of patients flagged as high risk were actually readmitted. Specificity averaged 0.59, ranging from 0.30 to 0.86, reflecting moderate ability to correctly identify patients who were not readmitted.

Despite testing more advanced machine-learning methods, their performance was similar to each other and to traditional approaches. Logistic regression, a commonly used statistical method, demonstrated the best calibration according to the Brier score, indicating that its predicted probabilities aligned more closely with observed outcomes than the other models.

The authors pointed to limitations in electronic medical record data as a key factor constraining predictive accuracy. They noted that many social determinants of health, such as housing stability and health literacy, are not routinely captured in the records used for modeling. As a result, important influences on readmission risk may be missing from the predictors available to both machine-learning models and logistic regression.

According to the study, incorporating structured information on social determinants at the time of intake could improve future algorithms. By systematically collecting variables that reflect patients’ living conditions and support systems, hospitals might be able to enhance risk prediction beyond what is achievable using clinical data alone. The findings suggest that data completeness, rather than model complexity, may be a central limitation for readmission prediction in this setting.