Machine Learning Characterization of Readmissions After Chronic Subdural Hematoma Hospitalizations

Readmission after chronic or subacute subdural hematoma (cSDH/sSDH) hospitalization is common, yet clinical attention and trial design have focused primarily on surgical recurrence. The full spectrum of readmission events, their clinical impact, and the heterogeneity of the affected patient population remain poorly understood. This study seeks to characterize the incidence, diversity, and outcomes of 90-day readmissions after cSDH/sSDH hospitalization and to identify patient phenotypes with distinct readmission risk profiles using machine learning-driven clustering.

This was a retrospective cohort study using the Nationwide Readmissions Database (2016-2022). Adults nonelectively hospitalized for cSDH/sSDH were included. The main outcome of interest was hospital readmission within 90 days, classified into readmission types. Readmission outcomes included length of stay, cost, in-hospital mortality, functional decline, new disability, and inability to return home. Machine learning-based phenotyping using Shapley Additive Explanations values from a multinomial gradient-boosted model and K-means clustering identified patient subgroups with divergent readmission patterns.

Of 22,387 patients (mean age 70.8 years; 29.6% female), 6,497 (29.0%) were readmitted within 90 days across diverse causes. Surgical SDH recurrence accounted for only 22.5% of readmissions, and fewer than half (44.0%) were primarily SDH-related. Non-SDH readmissions carried substantial clinical impact: infection readmissions had the highest mortality (9.6%), exceeding surgical SDH (2.9%) more than 3-fold, and the highest rate of new disability (45.5%) among patients initially discharged with routine self-care. Machine learning-driven phenotyping analysis identified 5 patient clusters with unique clinical characteristics and diverging readmission patterns: low acuity (39.8%), atrial fibrillation (16.9%), elderly/frail (16.6%), young/healthy (14.5%), and high acuity (12.1%). These phenotypes revealed marked patient heterogeneity, with each cluster exhibiting distinct readmission risk profiles.

Readmissions after cSDH/sSDH are diverse and predominantly nonsurgical, with non-SDH readmissions carrying equal or worse outcomes than surgical recurrence. Machine learning-based phenotyping uncovered substantial patient heterogeneity, highlighting the need for new therapeutic strategies, expanded clinical trial outcome targets beyond surgical recurrence, and comprehensive postdischarge care models tailored to distinct patient subgroups.