JMIR Publications

ABSTRACT

Background:

Discharge-planning nurses and case managers must identify which patients require intensive transitional care, but existing screening approaches often lack precision. Machine learning may support nursing prioritization if predictions are locally calibrated and linked to feasible nursing actions.

Objective:

This study aimed to develop and temporally validate a discharge-time readmission model and examine whether risk tiers could support nurse-led transitional care planning, including shared decision-making, long-term care linkage, and remote monitoring.

Methods:

We conducted a multicomponent observational study using 41,935 discharge episodes from 2017 to 2018. A 30-day all-cause readmission model was developed using 2017 discharges and temporally validated using 2018 discharges. Related public-benchmark validation used the UCI Diabetes 130 US Hospitals dataset with limited feature mapping. Separate institutional cohorts assessed the nursing actionability of shared decision-making for discharge destination (n=72), long-term care linkage (n=364), and remote monitoring among high-risk cardiopulmonary patients (n=43).

Results:

Random forest showed the strongest temporal validation performance (AUROC 0.786, 95% CI 0.772-0.797; AUPRC 0.333, 95% CI 0.312-0.353; Brier score 0.075, 95% CI 0.072-0.078), outperforming the existing rule-based high-risk flag (AUROC 0.501). Observed 30-day readmission increased from 3.1% in the low-risk tier to 45.6% in the very high-risk tier. In the public benchmark, logistic regression was most transportable (AUROC 0.567, 95% CI 0.561-0.572). Shared decision-making changed discharge direction in 41.7% of cases; long-term care began within 7 days in 90.3% of patients with valid service dates; and remote monitoring was associated with fewer 3-month hospital days, emergency visits, and readmissions.

Conclusions:

Discharge-time readmission prediction can provide a nursing-facing risk-stratification layer for targeted transitional care. Because transportability was limited and the action cohorts were observational, prospective silent-run testing, recalibration, workload monitoring, and evaluation of nurse-sensitive outcomes are needed before clinical activation.