Expanding care coordination in an integrated health system through causal machine learning

Abstract

Hospital readmission is a key quality metric, yet post-discharge interventions often yield variable results. In the first large-scale randomized evaluation of causal machine learning in a health system, we assessed whether a novel model (the Predicted Benefit Intervention (PBI) score) could identify lower-risk patients most likely to benefit from post-discharge care coordination within Kaiser Permanente Northern California (KPNC). From May to December 2022, 9959 low-risk patients at 19 KPNC hospitals were randomized to usual care or the Transitions Program, which included medication reconciliation, primary care follow-up scheduling, and weekly calls for 30 days. While 30-day readmissions declined in the intervention group (7.7% vs. 8.2%), the difference was not statistically significant. However, the observed-to-expected readmission ratio declined into randomization and remained low thereafter; this decline was statistically significant. This study demonstrates the feasibility of implementing causal machine learning at scale to improve targeting and resource allocation in care delivery.

Introduction

Over the past decade, hospital readmissions have grown into a major policy focus for hospitals, health systems, and payers. Growing recognition of the impact of readmissions has led to a range of initiatives aimed at preventing them, from targeted post-discharge care coordination programs implemented by individual hospitals to national efforts such as the Hospital Readmissions Reduction Program, which broadly incentivizes reducing readmissions. Despite these efforts, there is little consensus on which post-discharge interventions are most effective. While some programs have shown promise^{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12–13}, evidence suggests substantial variation in effectiveness across settings and target conditions^{2,5,6,9,14,15}.

Variation in the effectiveness of readmission prevention interventions may partly stem from how patients are identified for more focused post-discharge care. Many hospitals and health systems use score-based or algorithmic targeting, applying a scoring system, or statistical or machine learning models to predict readmission risk at discharge and prioritizing patients at highest predicted risk for intervention. However, it remains uncertain whether this approach reliably identifies patients whose readmission risk is modifiable—i.e., those at risk of preventable readmission¹⁶. To date, no hospitals or health systems have successfully implemented alternatives to risk-based algorithmic targeting at scale¹⁷. Given the limitations of risk-based algorithmic targeting for readmissions and other outcomes, causal machine learning has emerged as an alternative for generating more targeted predictions^16,17. Instead of predicting baseline risk, causal machine learning models estimate treatment response (or benefit) for a specific intervention.

Here, we describe a randomized evaluation at Kaiser Permanente of a quality improvement initiative to expand the Transitions Program, an intervention designed to prevent readmissions, to lower-risk patients using a causal machine learning model. In a prior study of high-risk discharges, the Transitions Program was associated with lower risks of post-discharge readmission and mortality¹. The data from this initial roll-out of the Transitions Program were then used to develop a causal machine learning model, the Predicted Benefit Intervention (PBI) score¹⁶. In this study, we evaluated the use of the PBI score to guide the expansion of the Transitions Program to lower-risk Kaiser Permanente patients, with the goal of reducing readmissions as measured by the Healthcare Effectiveness Data and Information Set (HEDIS) Plan All-Cause Readmissions metric.

Results

Characteristics of the randomized quality improvement evaluation cohort

During the randomized quality improvement evaluation period from May 19, 2022 to December 19, 2022, 52,754 patients were screened for eligibility prior to hospital discharge. Of these patients, 17,141 were eligible for randomization at discharge, of which 9959 were successfully randomized to either of two arms: (1) the usual post-discharge care pathway or (2) the Transitions Program intervention pathway, forming the primary intention-to-treat analysis cohort (Fig. 1). Baseline characteristics of these patients are presented in Table 1, while characteristics of otherwise eligible patients who were placed into alternate registries are shown in Supplementary Table 1. The median age in the intervention arm was 61 years; 55% were female. As indicated by their Transitions Support Level scores (median 13.1%, IQR 10.1–16.7%), these patients were at relatively low risk of readmission and/or death within 30 days of discharge.

Fig. 1 [Images not available. See PDF.]

Flow diagram of discharged KPNC patients screened for eligibility during the randomized period.

Table 1. Characteristics of the intention-to-treat cohort formed from KPNC patients successfully randomized during the randomized period

	Usual care arm n = 5057	Transitions Program arm n = 4902	Absolute standardized mean difference
Age	61 (45–74)	61 (44–74)	0.032
Male sex	2245 (44.4%)	2244 (45.8%)	0.028
Race and ethnicity			0.022
Asian	720 (14.2%)	707 (14.4%)
Black	548 (10.8%)	552 (11.3%)
Hispanic	1017 (20.1%)	994 (20.3%)
White	2463 (48.7%)	2369 (48.3%)
Other/unknown	309 (6.1%)	280 (5.7%)
Hospitalization not for COVID-19	5028 (99.4%)	4863 (99.2%)	0.027
Any ICU admission	459 (9.1%)	466 (9.5%)	0.015
Full code status	4518 (89.3%)	4394 (89.6%)	0.010
Index hospitalization length of stay, d	2.0 (1–4)	2 (1–4)	0.022
SNP/Medi-Cal	660 (13.1%)	681 (13.9%)	0.025
SNP	380 (7.5%)	359 (7.3%)	0.0073
Medi-Cal	312 (6.2%)	359 (7.3%)	0.046
LAPS2 score at admission	71 (52–92)	71 (52–91)	0.019
LAPS2 score at discharge	44 (28–63)	44 (29–64)	0.0079
COPS2 score	35 (13–64)	34 (13–65)	0.0072
TSL score	13.2 (10.1–16.8)	13.0 (10.1–16.7)	0.030
Prior Transitions engagement	1119 (22.1%)	1100 (22.4%)	0.0075
Prior Transitions status			0.054
No prior Transitions engagement	3938 (77.9%)	3802 (77.6%)
Referral review closure	210 (4.2%)	169 (3.4%)
Unable to reach	75 (1.5%)	94 (1.9%)
Declined	66 (1.3%)	58 (1.2%)
Enrolled	768 (15.2%)	779 (15.9%)

The COmorbidity Point Score, version 2 (COPS2)²⁷ is assigned every month to all adults with a Kaiser Permanente Northern California medical record number based on diagnoses incurred in the preceding 12 months. Range is from 0 to 1010; higher scores indicate worse mortality risk. The univariate relationship between the COPS2 and 1-year mortality is as follows: 0–39, 0.3%; 40–64, 5.3%; ≥65, 17.2%.

The admission Laboratory-based Acute Physiology Score, version 2 (LAPS 2)²⁸ is assigned on the basis of a patient’s worst vital signs, pulse oximetry, neurologic status, and 16 laboratory test results in the 72 h preceding hospital entry. The univariate relationship of an admission LAPS2 with 30-day mortality is as follows: 0–59, 1.0%; 60–109, 5.0%; 110+, 13.7%. After age, sex, diagnosis, and comorbid conditions are controlled for, the adjusted odds ratio for inpatient mortality for an increase in LAPS2 of 5 points is 1.13. Note that this hospital score is not assigned to outpatient visits.

For continuous variables, the median and IQR (in parentheses) are reported. CHF chronic heart failure, ICU intensive care unit, SNP Special Needs Program, LAPS2 score Laboratory-based Acute Physiology score, version 2²⁸, COPS2 score Comorbidity Burden Point Score, version 2, TSL score Transitions Support Level score.

Outcomes during the randomized quality improvement period

In the primary intention-to-treat analysis, 415 of 5057 patients (8.2%) in the usual care arm experienced a non-elective rehospitalization within 30 days, compared with 378 of 4902 (7.7%) patients in the intervention arm. The unadjusted and adjusted risk ratios for 30-day re-hospitalization were 0.94 (95% CI 0.82–1.07) and 0.94 (95% CI 0.81–1.06), respectively. There were three deaths within 30 days of discharge in the usual care arm, compared with one in the intervention arm; thus, 418 patients in the usual care arm (8.3%) experienced the composite outcome of non-elective rehospitalization and/or death, compared with 379 patients (7.7%) in the intervention arm. The unadjusted and adjusted risk ratios for this composite outcome were 0.94 (95% CI 0.82–1.07) and 0.93 (95% CI 0.81–1.05), respectively(Table 2).

Table 2. Outcomes in the intention-to-treat cohort during the randomized period

Outcome	Usual care arm n = 5057	Transitions Program arm n = 4902	Unadjusted risk ratio (95% CI)	Adjusted risk ratio (95% CI)
HEDIS-reportable NEH	415 (8.2%)	378 (7.7%)	0.94 (0.82–1.07)	0.94 (0.81–1.06)
Post-discharge mortality	3	1	–	–
HEDIS-reportable NEH or mortality	418 (8.3%)	379 (7.7%)	0.94 (0.82–1.07)	0.93 (0.81–1.05)

All outcomes are measured at 30 days following discharge. Abbreviations:HEDIS Healthcare Effectiveness Data and Information Set, NEH non-elective rehospitalization.

In the per-protocol analysis accounting for noncompliance using the distillation method, 285 of 3702 patients (7.7%) in the usual care arm experienced a re-hospitalization within 30 days, compared with 261 of 3579 (7.3%) in the intervention arm. The unadjusted and adjusted risk ratios for 30-day re-hospitalization and/or death were 0.95 (95% CI 0.81–1.11) and 0.95 (95% CI 0.79–1.09), respectively. The characteristics of both arms following distillation are presented in Supplementary Table 2. There were two deaths in the distilled usual care arm and none in the distilled intervention arm; thus, 287 patients in the distilled usual care arm (8.3%) experienced the composite outcome of non-elective rehospitalization and/or death, compared with 261 patients (7.7%) in the distilled intervention arm. The unadjusted and adjusted risk ratios for this composite outcome were 0.94 (95% CI 0.80–1.11) and 0.93 (95% CI 0.78–1.09), respectively (Table 3).

Table 3. Outcomes in the distilled per-protocol cohort during the randomized period

Outcome	Usual care arm n = 3702	Transitions Program arm n = 3579	Unadjusted risk ratio (95% CI)	Adjusted risk ratio (95% CI)
HEDIS-reportable NEH	285 (7.7%)	261 (7.3%)	0.95 (0.81–1.11)	0.95 (0.79–1.09)
Post-discharge mortality	2	0	–	–
HEDIS-reportable NEH or mortality	287 (8.3%)	261 (7.7%)	0.94 (0.80–1.11)	0.93 (0.78–1.09)

All outcomes are measured at 30 days following discharge. Refer tothe "Methods" section for details on the distillation method and its implementation. Abbreviations:HEDIS Healthcare Effectiveness Data and Information Set, NEH non-elective rehospitalization.

Observed outcomes before and after randomization

In the pre-randomization period, which spanned from March 4, 2019 to April 25, 2022 and included 46,058 TSL low-risk patients who would otherwise have been eligible for randomization and thus are comparable to the intention-to-treat randomized cohort above, 4113 patients were referred to the Transitions Program at discharge. Of these patients, 515 (12.5%) experienced a re-hospitalization and/or death within 30 days. Among 41,945 patients who were not selected for outreach, 4171 (9.9%) experienced the same outcome, resulting in unadjusted and adjusted risk ratios for 30-day re-hospitalization and/or death of 1.26 (95% CI 1.17–1.37) and 1.15 (95% CI 1.04–1.27), respectively. These estimated risk ratios were also not significantly different from those in the distilled pre-randomization cohort. (Supplementary Tables 3 and 4.)

Furthermore, the observed-to-expected ratio of 30-day re-hospitalization was 0.97 (95% CI 0.94–1.00) during the pre-randomization period, compared to 0.79 (95% CI 0.74–0.85) during the randomized period, a statistically significant difference. The reduction in observed-to-expected ratio from the pre-randomized period to the randomized period was sustained in the post-randomization period (December 19, 2022, to June 30, 2023) during which time this ratio was 0.81 (95% CI 0.76–0.87). The time course of this observed-to-expected ratio is shown in Fig. 2.

Fig. 2 [Images not available. See PDF.]

Monthly observed-to-expected ratios of 30-day rehospitalization over all three periods (including the randomized period demarcated by the dashed lines), stratified by Transitions Program referral.

Discussion

In this randomized quality improvement evaluation, we implemented and evaluated an expanded version of the Transitions Program, a post-discharge care coordination intervention, among hospitalized patients identified as low risk for readmission but with high predicted benefit using a causal machine learning model (the PBI score). We observed a modest reduction in 30-day rehospitalization rates during the randomized period, which, although not statistically significant, was consistent in magnitude with prior evaluations of the Transitions Program¹. Notably, the observed-to-expected ratio of 30-day rehospitalization declined significantly from the pre-randomization to the randomized period, with this improvement sustained into the post-randomization period, suggesting the durability of the potential impact of benefit-based targeting.

To our knowledge, this study represents the first randomized evaluation of a care coordination intervention targeted using a causal machine learning approach. Unlike traditional risk models, which assume that individuals at highest risk are most likely to benefit, causal machine learning methods estimate individual treatment effects directly, aiming to identify those most likely to respond to an intervention. While risk prediction models are well-established and easier to develop and validate retrospectively, growing evidence suggests that predicted risk is not always a reliable proxy for treatment response^{18, 19–20}, including in the context of readmission prevention¹⁶. Despite this, causal machine learning models require retrospective data on both treated and untreated populations, which may limit their feasibility in some settings.

Our study has several strengths and limitations. First, to our knowledge, it is the first large-scale implementation and evaluation of a clinical intervention targeted using causal machine learning. Second, the study was conducted within a large, integrated health care system, allowing us to enroll a diverse population and achieve near-complete capture of the primary outcome of readmission. This enhances the generalizability of our findings across a wide range of demographic and clinical subgroups. While ascertainment of mortality may have been incomplete due to delays in updating death records, we have no reason to believe such delays differed systematically between trial arms or introduced bias. Moreover, because the study focused on a lower-risk hospitalized population, the expected number of deaths was small. Third, unlike many randomized evaluations that assess outcomes only during the intervention period, our study integrated findings from randomized data with observations from pre- and post-randomization periods, providing additional evidence that reductions in readmissions may have been sustained as the intervention was more broadly implemented.

One major limitation of our study is the relatively high rate of cross-over between trial arms, driven by referral review closures by case managers and differences in patients’ willingness to enroll in the Transitions Program. These manual closures may reflect a form of “soft” algorithm aversion²¹ to the extent that case managers were aware of algorithmic recommendations. In either case, these closures were mainly driven by redundancy of the Transitions Program with existing services (e.g., home health care) or by clinical judgment that the Transitions Program was not the most appropriate care pathway for patients with specialized needs (e.g., active alcohol use disorder). However, sensitivity analyses using the distillation method to account for cross-over yielded results consistent with our primary analysis. Moreover, because these cross-over mechanisms appeared to affect both trial arms similarly, we do not believe they introduced bias into our effect estimates. Rather, their primary impact was likely through reducing statistical power. Indeed, another key limitation was that the randomized design had lower power than originally planned, largely due to this cross-over. In addition, the “siphoning off” of eligible patients into registries upstream of randomization (see Fig. 1 and Supplementary Table 1) further reduced power. Together, these limitations highlight the challenges of integrating randomized evaluations of algorithm-based interventions into existing care processes that retain a “human in the loop” with oversight of final treatment decisions.

Finally, while our evaluation focused on HEDIS-reportable readmissions, other categories of readmissions among low-risk patients may also represent important opportunities for tailored care coordination. These include readmissions related to perinatal conditions or experienced by high utilizers not already covered by other programs. Currently, many of these patients are already served by parallel programs within KPNC, including perinatal care pathways and high-risk outreach through the existing high-risk segment of the Transitions Program. In addition, we chose to focus on patients meeting HEDIS criteria to align our evaluation with the target population of the PBI model. Altogether, the exclusion of these patients from our evaluation reflects both the current state of post-discharge coordination at KPNC as well as the design of the PBI model. Future studies could explore the effectiveness of new benefit-based targeting strategies, tailored to the needs of these specific populations.

In conclusion, expanding the Kaiser Permanente Transitions Program by targeting the lower-risk patients most likely to benefit from enrollment using a causal machine learning approach yielded reductions in readmissions as observed during a randomized quality improvement period. While these reductions were not statistically significant, their magnitude was consistent with observed associations in a prior study. However, incorporating observational data from before and after randomization showed that the Transitions Program was associated with significant and sustained reductions in the observed-to-expected ratio of readmissions following randomization.

Methods

Background and study design

This report describes a randomized evaluation of a quality improvement initiative conducted in 19 of 21 Kaiser Permanente Northern California (KPNC) hospitals to assess a targeted expansion of the Transitions Program, a post-discharge care pathway, to hospitalized patients with low predicted readmission risk. This expansion was guided by a causal machine learning model designed to predict response to the Transitions Program. As originally implemented, the Transitions Program is a care coordination intervention offered to eligible hospitalized patients immediately after discharge based on their Transitions Support Level (TSL) score, which estimates the risk of non-elective re-hospitalization and/or death within 30 days of discharge²². Before this study, only high-risk patients (TSL-predicted risk >25%) were routinely referred to the program¹.

Following the observed effectiveness of the Transitions Program in high-risk patients, the decision was made to expand it to hospitalized patients discharged at lower risk (TSL < 25%). Given the much larger volume of lower-risk patients, we prioritized expansion based on individualized predicted benefit (i.e., estimated treatment effect) rather than predicted risk, using a causal machine learning approach informed by our prior work¹⁶. To facilitate implementation, we developed and deployed a statistical model (the Predicted Benefit Intervention (PBI) score) based on the T-learner meta-algorithm²³, using logistic regression as the base learner and trained on data from before and after Transitions Program implementation. The treatment and control models included the same covariates as those used in the TSL score²², and were tuned separately via cross-validation to optimize the binary log-loss. As part of the cross-validation process, we varied the length of the pre/post-implementation window, defined relative to the Transitions Program rollout date at each medical center, to identify the smallest window (~300 days on either side) that maintained good calibration and discrimination within each treatment group. This approach traded off sample size for temporal proximity to rollout, while preserving the calibration of the outcome models that were used to estimate treatment effects.

Eligibility and data collection

The data used in this study derive from the TSL-low risk population followed over three non-overlapping time periods between March 4, 2019, and June 30, 2023.

First, the pre-randomization period (March 4, 2019 to April 25, 2022) included all hospitalized KPNC patients who survived to discharge. During this period, no systematic post-discharge outreach was conducted in the TSL low-risk group, except for manual selection by case managers beginning in May 2021. These discharges were retrospectively scored to generate PBI scores for these low-risk hospitalized patients, allowing us to identify those who would have been eligible for the PBI pathway — and thus for randomization (see Supplementary Fig. 1)—had their discharge occurred during the subsequent randomized period, described below.

Second, the randomized quality improvement period (May 19 to December 19, 2022) followed a washout period (April 26 to May 18, 2022) and is the primary focus of this report. During this phase, both TSL and PBI scores were used to identify eligible TSL low-risk patients, who were then randomized to receive either the Transitions Program intervention or usual post-discharge care and were followed for 30 days post-discharge.

Hospitalized patients were eligible for randomization following discharge if they met the following criteria. First, they were required to be at low risk of readmission, defined by a TSL score below 25%. Second, they were required to have an estimated PBI score below a predetermined hospital-level threshold. These thresholds were set to align discharge volume with outreach capacity within each Northern California service area. Patients meeting both criteria were classified as predicted low-risk/high-benefit and were eligible for randomization.

Randomization was implemented at the discharge encounter level within the KP HealthConnect electronic health record (EHR) system. Each eligible patient had an equal probability of being assigned to either the intervention or usual care group. Some patients who met the criteria above were excluded from outreach if they qualified for pre-existing specialized registries (Fig. 1). Furthermore, while all eligible patients were randomized, case managers retained discretion to override assignments based on clinical and operational considerations. This included offering Transitions Program enrollment to patients in the control arm or, conversely, closing referrals for patients assigned to the intervention arm through entering a referral review closure (RRC). Examples of common free-text reasons given for RRCs include “with active Home Health orders” and “exclusionary criteria of active alcohol abuse”.

Finally, in the post-randomization period (December 19, 2022 to June 30, 2023), randomization was discontinued, and all patients meeting eligibility criteria were enrolled in the PBI workflow. A summary of patient flow across all three study periods is provided in Supplementary Fig. 1.

Intervention

During the randomized quality improvement period, patients randomized to the intervention arm were referred to the Transitions Program following discharge. Patients who accepted an active referral were enrolled and followed by the program for 30 days, receiving the same outreach as high-risk patients. This included post-discharge care coordination, medication reconciliation, and weekly telephone consultations with a Transitions Program case manager, following existing protocols¹.

In the post-randomization period, all eligible patients received the same Transitions Program intervention as did those enrolled during the randomized period.

Outcome measures

The three primary outcomes in this study included the following, two of which were based on the HEDIS criteria²⁴: (1) HEDIS-reportable re-hospitalization, defined as non-elective re-hospitalization within 30 days of dischargemeeting HEDIS criteria; (2) 30-day post-discharge mortality; and (3) a composite outcome of 30-day HEDIS-reportable re-hospitalization or mortality.

HEDIS criteria exclude patients with frequent hospitalizations (≥4 admissions in the measurement year for Medicare and Medicaid beneficiaries and ≥3 for other payers) as well as those with a primary diagnosis related to pregnancy or perinatal conditions. Our analysis focused on HEDIS-eligible patients to align with operational and methodological considerations: HEDIS-reportable readmissions are the primary outcome metric used for quality measurement and outcome reporting, and reflecting this, the base models for the PBI score were trained using data from HEDIS-eligible patients. Because real-time determination of HEDIS eligibility was not feasible, analytic cohorts were formed retrospectively by applying these criteria to successfully randomized patients.

Statistical analysis

The randomized quality improvement evaluation of benefit-based targeting using the PBI score was designed to enroll ~9000 patients during the second (randomized) period. This sample size provided 80% power to detect a 10% relative risk reduction between arms for the primary outcome of 30-day re-hospitalization and/or death. The expected effect size was based our prior study evaluating the Transitions Program in high-risk hospitalized patients¹.

Our primary analysis assessed 30-day re-hospitalization in the intention-to-treat population, which included all randomized patients regardless of their final treatment assignment. Cross-over between trial arms could occur through three primary mechanisms: (1) enrollment declination, where patients randomized to the intervention arm declined to enroll in the Transitions Program; (2) RRC, where case managers closed a referral for an intervention arm patient, precluding Transitions Program outreach; and (3) ad hoc referrals, where patients randomized to usual care were subsequently referred to the Transitions Program, though this occurred in only four discharges.

Because we anticipated a high level of cross-over between trial arms due to enrollment declinations, RRCs for intervention arm patients, and ad hoc referrals for usual care arm patients, we planned a sensitivity analysis using the distillation method²⁵. This approach, a form of per-protocol analysis, first builds a model to predict cross-over (defined as enrollment declination or RRC) within the intervention arm, akin to a propensity score. This model is then applied to exclude patients from both arms with predicted probabilities of enrollment below a specified threshold, effectively “distilling” the data into subsets of patients most likely to enroll. The primary estimator is then re-applied to these distilled data. For comparison, we applied the same approach to the observational, pre-randomization data to construct an observational analog of the per-protocol cohort.

Augmented inverse propensity-weighted logistic regression was used to adjust for baseline covariates at discharge in all analyses. These models adjusted for age, sex, race and ethnicity, KP membership, Medi-Cal or Special Needs Plan status, TSL and PBI scores at discharge, whether the hospital admission originated in the emergency department, discharging KP medical center, prior engagement with the Transitions Program, and diagnoses of COVID-19, sepsis, or chronic heart failure during hospitalization. Additionally, we estimated observed-to-expected ratios for 30-day non-elective re-hospitalization using a version of the TSL score to compute the expected number of re-hospitalizations within each period.

This study was approved as a quality improvement evaluation by the KPNC Research Determination Office. In addition, where applicable, the reporting of this study conforms to the guidelines recommended by the CONSORT statement²⁶.

Acknowledgements

This study was funded by Kaiser Foundation Health Plan and Kaiser Foundation Hospitals. The authors are grateful to the KP HealthConnect team as well as to the Transitions Program case management and clinical operations teams across Kaiser Permanente Northern California for their partnership in implementing the PBI algorithm and associated intervention.

Author contributions

B.J.M. co-led study design, oversaw data analysis, interpreted results, and drafted the manuscript. C.P. contributed to data management and analysis, visualization, and manuscript preparation. P.K. co-led study design, designed the randomized evaluation, and supervised statistical modeling. G.J.E. contributed to study conception, implementation planning, and manuscript revision. L.C.M. provided clinical oversight and helped interpret findings. M.C.D. and J.R.S. contributed to the methods section. J.D.G. provided statistical support and data extraction. M.D.F., M.C.D., and J.R.S. liaised between the study team and frontline clinical teams, supported program implementation, and assisted in coordination of operational workflows; they also contributed to the manuscript. V.X.L. supervised all aspects of the study, including design, analysis, and writing. All authors reviewed and approved the final manuscript.

Data availability

The data that support the findings of this study are not publicly available due to patient privacy and institutional data governance policies. Access to the data is restricted and governed by Kaiser Permanente. Researchers with specific inquiries may contact the corresponding author, and requests will be considered on a case-by-case basis in accordance with institutional policies and applicable approvals.

Code availability

Analytic code used in this study is available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests relevant to this work.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41746-025-01925-3.

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1. Marafino, BJ et al. Evaluation of an intervention targeted with predictive analytics to prevent readmissions in an integrated health system: observational study. BMJ; 2021; 374, n1747. [DOI: https://dx.doi.org/10.1136/bmj.n1747] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34380667][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8356037]

2. Bilicki, D. J. Outpatient follow-up visits to reduce 30-day all-cause readmissions for heart failure, copd, myocardial infarction, and stroke: a systematic review and meta-analysis. Prev. Chronic Dis. 21, E74 (2024).

3. Tyler, N et al. Transitional care interventions from hospital to community to reduce health care use and improve patient outcomes: a systematic review and network meta-analysis. JAMA Netw. Open; 2023; 6, e2344825. [DOI: https://dx.doi.org/10.1001/jamanetworkopen.2023.44825] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38032642][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10690480]

4. Gonçalves-Bradley, DC; Lannin, NA; Clemson, L; Cameron, ID; Shepperd, S. Discharge planning from hospital. Cochrane Database Syst. Rev.; 2022; 2, CD000313. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35199849]

5. Chauhan, U; McAlister, FA. Comparison of mortality and hospital readmissions among patients receiving virtual ward transitional care vs usual postdischarge care: a systematic review and meta-analysis. JAMA Netw. Open; 2022; 5, e2219113. [DOI: https://dx.doi.org/10.1001/jamanetworkopen.2022.19113] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35763296][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9240908]

6. Bamforth, RJ et al. Strategies to prevent hospital readmission and death in patients with chronic heart failure, chronic obstructive pulmonary disease, and chronic kidney disease: a systematic review and meta-analysis. PLoS ONE; 2021; 16, e0249542.1:CAS:528:DC%2BB3MXhtVSks7nK [DOI: https://dx.doi.org/10.1371/journal.pone.0249542] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33886582][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8062060]

7. Becker, C et al. Interventions to improve communication at hospital discharge and rates of readmission: a systematic review and meta-analysis. JAMA Netw. Open; 2021; 4, e2119346. [DOI: https://dx.doi.org/10.1001/jamanetworkopen.2021.19346] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34448868][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8397933]

8. Dautzenberg, L et al. Medication review interventions to reduce hospital readmissions in older people. J. Am. Geriatr. Soc.; 2021; 69, pp. 1646-1658. [DOI: https://dx.doi.org/10.1111/jgs.17041] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33576506][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8247962]

9. Penney, LS et al. Interventions to reduce readmissions: can complex adaptive system theory explain the heterogeneity in effectiveness? A systematic review. BMC Health Serv. Res; 2018; 18, [DOI: https://dx.doi.org/10.1186/s12913-018-3712-7] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30477576][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6260570]894.

10. Uminski, K et al. Effect of post-discharge virtual wards on improving outcomes in heart failure and non-heart failure populations: a systematic review and meta-analysis. PLoS ONE; 2018; 13, e0196114. [DOI: https://dx.doi.org/10.1371/journal.pone.0196114] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29708997][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5927407]

11. Braet, A; Weltens, C; Sermeus, W. Effectiveness of discharge interventions from hospital to home on hospital readmissions: a systematic review. JBI Database Syst. Rev. Implement Rep.; 2016; 14, pp. 106-173. [DOI: https://dx.doi.org/10.11124/jbisrir-2016-2381]

12. Vedel, I; Khanassov, V. Transitional care for patients with congestive heart failure: a systematic review and meta-analysis. Ann. Fam. Med.; 2015; 13, pp. 562-571. [DOI: https://dx.doi.org/10.1370/afm.1844] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/26553896][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4639382]

13. Leppin, AL et al. Preventing 30-day hospital readmissions: a systematic review and meta-analysis of randomized trials. JAMA Intern. Med.; 2014; 174, pp. 1095-1107. [DOI: https://dx.doi.org/10.1001/jamainternmed.2014.1608] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/24820131][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4249925]

14. Donzé, J et al. Effects of a multimodal transitional care intervention in patients at high risk of readmission: the TARGET-READ randomized clinical trial. JAMA Intern. Med.; 2023; 183, pp. 658-668. [DOI: https://dx.doi.org/10.1001/jamainternmed.2023.0791] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37126338][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10152373]

15. Wachter, RM. The disappointing impact of interventions to prevent hospital readmissions. JAMA Intern. Med.; 2023; 183, pp. 668-669. [DOI: https://dx.doi.org/10.1001/jamainternmed.2023.0804] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/37126339]

16. Marafino, BJ; Schuler, A; Liu, VX; Escobar, GJ; Baiocchi, M. Predicting preventable hospital readmissions with causal machine learning. Health Serv. Res.; 2020; 55, pp. 993-1002. [DOI: https://dx.doi.org/10.1111/1475-6773.13586] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33125706][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7704477]

17. Feuerriegel, S et al. Causal machine learning for predicting treatment outcomes. Nat. Med.; 2024; 30, pp. 958-968.1:CAS:528:DC%2BB2cXovF2jtLs%3D [DOI: https://dx.doi.org/10.1038/s41591-024-02902-1] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/38641741]

18. Kent, DM et al. Risk and treatment effect heterogeneity: re-analysis of individual participant data from 32 large clinical trials. Int. J. Epidemiol.; 2016; 45, pp. 2075-2088. [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27375287][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5841614]

19. Ascarza, E. Retention futility: targeting high-risk customers might be ineffective. J. Mark. Res. 55, (2018). Accessed 8 May 2024. https://journals.sagepub.com/doi/abs/10.1509/jmr.16.0163.

20. Fernández-Loría, C; Provost, F. Causal classification: treatment effect estimation vs. outcome prediction. J. Mach. Learn Res; 2022; 23, pp. 1-35.

21. Sunstein, C. R. & Gaffe, J. An anatomy of algorithm aversion. Published online June 14, https://doi.org/10.2139/ssrn.4865492 (2024).

22. Escobar, GJ et al. Nonelective rehospitalizations and postdischarge mortality: predictive models suitable for use in real time. Med. Care; 2015; 53, pp. 916-923. [DOI: https://dx.doi.org/10.1097/MLR.0000000000000435] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/26465120][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4605276]

23. Künzel, SR; Sekhon, JS; Bickel, PJ; Yu, B. Metalearners for estimating heterogeneous treatment effects using machine learning. Proc. Natl. Acad. Sci. USA; 2019; 116, pp. 4156-4165. [DOI: https://dx.doi.org/10.1073/pnas.1804597116] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30770453][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6410831]

24. National Committee for Quality Assurance. HEDIS, 2, Technical Specifications. Published online 2019 https://store.ncqa.org/hedis-2019-volume-2-epub.html (NCQA, 2019).

25. Adams, JL; Davis, AC; Schneider, EC; Hull, MM; McGlynn, EA. The distillation method: a novel approach for analyzing randomized trials when exposure to the intervention is diluted. Health Serv. Res; 2022; 57, pp. 1361-1369. [DOI: https://dx.doi.org/10.1111/1475-6773.14014] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35752926][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9643092]

26. Schulz, KF; Altman, DG; Moher, D. CONSORT Group. CONSORT 2010 statement: updated guidelines for reporting parallel group randomised trials. BMJ; 2010; 340, c332. [DOI: https://dx.doi.org/10.1136/bmj.c332] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/20332509][PubMedCentral: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2844940]

27. Escobar, GJ; Gardner, MN; Greene, JD; Draper, D; Kipnis, P. Risk-adjusting hospital mortality using a comprehensive electronic record in an integrated health care delivery system. Med. Care; 2013; 51, pp. 446-453. [DOI: https://dx.doi.org/10.1097/MLR.0b013e3182881c8e] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23579354]

28. Escobar, GJ et al. Risk-adjusting hospital inpatient mortality using automated inpatient, outpatient, and laboratory databases. Med. Care; 2008; 46, pp. 232-239. [DOI: https://dx.doi.org/10.1097/MLR.0b013e3181589bb6] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/18388836]