Health systems face persistent pressure to reduce hospital readmissions, yet outcomes from post-discharge programmes vary across settings and target conditions. An integrated approach evaluated whether causal machine learning could help allocate care coordination resources to patients most likely to benefit despite being at lower baseline risk. A randomised quality improvement period expanded the Transitions Program to such patients, guided by a Predicted Benefit Intervention (PBI) score embedded in the electronic health record. Outcomes focused on 30-day non-elective rehospitalisation, mortality and a composite measure aligned with Healthcare Effectiveness Data and Information Set (HEDIS) reporting. The evaluation combined randomised results with pre- and post-randomisation observations to examine effect size, durability and operational feasibility at scale.
Randomised Evaluation Across 19 Hospitals
From 19 May to 19 December 2022, 52,754 hospitalised patients were screened before discharge at 19 of 21 hospitals. Of these, 17,141 met eligibility for randomisation and 9,959 were successfully randomised in the intention-to-treat cohort. Eligibility required a Transitions Support Level (TSL) risk score below 25% and a PBI score meeting hospital-level thresholds calibrated to outreach capacity. The PBI score was developed with a T-learner meta-algorithm using logistic regression base learners, trained on data before and after Transitions Program implementation and tuned via cross-validation to balance calibration and discrimination within treatment and control groups.
Randomisation occurred at the discharge encounter level in the electronic health record, allocating patients to either usual post-discharge care or referral to the Transitions Program. The intervention comprised post-discharge care coordination, medication reconciliation and weekly telephone consultations with a case manager for 30 days, mirroring support delivered to high-risk patients. Clinicians could override assignments based on clinical and operational judgement, for example when services overlapped with existing home health care or when specialised needs made the pathway unsuitable. These referral review closures and patient declinations produced cross-over between arms that was anticipated in the design.
Baseline characteristics reflected a lower-risk population. In the intervention arm the median age was 61 years and 55% were female. Median TSL was 13.1% with an interquartile range of 10.1–16.7%, indicating comparatively low predicted risk of 30-day rehospitalisation or death. Additional characteristics, such as Laboratory-based Acute Physiology Score (version 2) and comorbidity burden scores, were similar between arms, supporting comparability for intention-to-treat analyses.
Modest Outcome Shifts and Sustained Ratio Declines
In intention-to-treat analyses, 30-day HEDIS-reportable non-elective rehospitalisation occurred in 8.2% of usual care patients (415 of 5,057) and 7.7% of intervention patients (378 of 4,902). The unadjusted and adjusted risk ratios were both 0.94 with 95% confidence intervals of 0.82–1.07 and 0.81–1.06, respectively. There were three deaths in the usual care arm and one in the intervention arm, yielding composite outcome rates of 8.3% versus 7.7% and corresponding risk ratios of 0.94 unadjusted and 0.93 adjusted with confidence intervals spanning unity.
Given cross-over, a per-protocol sensitivity analysis applied a distillation method to identify subgroups most likely to enrol in the Transitions Program. In these distilled cohorts, 30-day rehospitalisation was 7.7% in usual care (285 of 3,702) and 7.3% in intervention (261 of 3,579), with unadjusted and adjusted risk ratios of 0.95 and 0.95 (95% confidence intervals 0.81–1.11 and 0.79–1.09). The composite of rehospitalisation or mortality was 8.3% versus 7.7% with risk ratios of 0.94 unadjusted and 0.93 adjusted (95% confidence intervals 0.80–1.11 and 0.78–1.09). Point estimates were consistent in magnitude with prior evaluations of the Transitions Program, although not statistically significant within the power available.
Must Read: Enhancing Hospital Discharge to Prevent Readmissions
The evaluation also examined observed-to-expected ratios for 30-day rehospitalisation using a TSL-based expected outcome model. In the pre-randomisation period (4 March 2019 to 25 April 2022), the ratio was 0.97 with a 95% confidence interval of 0.94–1.00 among comparable low-risk patients. During randomisation, the ratio decreased to 0.79 (0.74–0.85), representing a statistically significant decline. This improvement persisted after randomisation ended, with a ratio of 0.81 (0.76–0.87) from 19 December 2022 to 30 June 2023. Observational comparisons from the pre-randomisation phase showed higher composite outcome rates among manually selected patients than among those not selected for outreach, aligning with the rationale for benefit-based targeting rather than risk-only strategies.
Implementation, Strengths and Constraints
The expansion targeted a larger pool of lower-risk discharges, necessitating prioritisation by predicted benefit rather than predicted risk alone. The PBI score incorporated the same covariates as the TSL risk model but was trained separately within treated and untreated groups to estimate treatment response. Cross-validation selected a window of approximately 300 days before and after programme rollout for model training, balancing temporal relevance and sample size while maintaining calibration. Randomisation integrated with routine care processes, recognising that case managers would retain discretion and that patients might decline outreach.
Strengths included scale, integration and outcome capture. The initiative spanned 19 hospitals in a unified system with near-complete ascertainment of 30-day readmission, enhancing generalisability across demographic and clinical subgroups. The design combined randomised evidence with pre- and post-randomisation observations, providing context on durability as the workflow transitioned from trial allocation to full implementation for eligible patients. The focus on HEDIS-reportable outcomes aligned with operational quality metrics and with the data used to train base models underlying the PBI score.
Constraints reflected real-world complexity. Cross-over due to referral closures and patient preferences reduced statistical power and diluted exposure differences between arms. Upstream assignment to specialised registries before randomisation further limited enrolment. Mortality ascertainment could be affected by delays in updating death records, although such delays were not expected to differ systematically between arms, and deaths were infrequent in this lower-risk cohort. The approach depends on retrospective data from both treated and untreated populations, which may limit feasibility in settings without comparable historical cohorts or consistent programme deployment. Additionally, the evaluation deliberately excluded categories outside HEDIS scope, such as perinatal readmissions or high utilisers managed by other programmes, to align with model design and operational priorities.
Targeting lower-risk hospital discharges for care coordination based on predicted benefit was feasible at scale and associated with a sustained reduction in observed-to-expected 30-day rehospitalisation ratios as the workflow matured from randomised allocation to routine use. Although absolute differences in rehospitalisation and composite outcomes were modest and not statistically significant within the randomised period, effect sizes were consistent with previous observations of the Transitions Program. The experience illustrates how causal machine learning can complement established risk models to direct finite outreach resources, embed targeting within everyday clinical operations and monitor impact using routine quality metrics without disrupting existing human oversight.
Source: npj digital medicine
Image Credit: Freepik
