Development and internal validation of a 30-day readmission risk model for older adults using Swiss electronic health record data: a retrospective cohort study Authors: Laura Maria Steiner, Sandra M G Zwakhalen, Loris Bonetti, Sabine Hahn Published: 2026 ABSTRACT AbstractBackgroundUnplanned 30-day readmissions in discharged older adults often reflect unmet needs and fragmented care transitions. Existing prediction models frequently lack geriatric specificity, clinical interpretability, and applicability to decentralised European healthcare systems. This study aimed to develop a clinically interpretable risk model for older adults using routine discharge data. Findings may provide a basis for targeted interventions.MethodsWe conducted a retrospective cohort study using electronic health record (EHR) data from a Swiss multi-site hospital network. Adults aged ≥ 65 years discharged alive in 2024 were eligible, excluding planned readmission, oncology/palliative cases, in-hospital deaths, and records with missing predictors. Of 15,635 discharges, 12,814 met inclusion criteria. 9,429 (60.3%) were retained for complete-case analysis. A multivariable logistic regression model with backward selection predicted unplanned 30-day readmission. Model performance was evaluated by AUC, calibration, decision curve analysis (DCA), and bootstrap validation, with a Fine-Gray model addressing competing risk of mortality.ResultsAmong 9,429 patients (mean age 78.0 years; 45.5% female), 1,845 (19.6%) had an unplanned 30-day readmission. Independent predictors were polypharmacy (OR 1.83), home care (OR 1.61), male sex (OR 1.42), low fall risk (vs. high; OR 1.41), moderate care intensity (vs. high; OR 1.30), comorbidity count (OR 1.12/diagnosis), younger age (OR 0.99/year), and shorter stay (OR 0.98/day). The model showed an AUC of 0.716 (95% CI 0.706–0.726), but calibration was poor, with risk overestimation at low predicted probabilities. At a 10% threshold, sensitivity was 7.5%, specificity 98.1%, positive predictive value 49.5%, and negative predictive value 81.4%. Compared with LACE index (AUC 0.707, 95% CI 0.694–0.720), performance improved modestly but significantly (p < 0.01).ConclusionsNearly one in five older adults was readmitted within 30 days, underscoring the burden of rehospitalisation and its link to quality-of-care. The model identified clinically relevant, EHR-available predictors, including medical complexity and care dependency. Despite acceptable discrimination, calibration was suboptimal and sensitivity low, limiting immediate clinical application. Future refinement and external validation may enhance utility. This model may provide a pragmatic starting point for risk-stratified discharge planning research and guiding transitional care interventions. Trial registrationStudy protocol was registered on the Open Science Framework (OSF) (https://osf.io/vkte2/; 10 May 2025).Supplementary InformationThe online version contains supplementary material available at 10.1186/s12877-026-07468-w.Keywords: Hospital readmission, Older adults, Prediction model, Electronic health records, Risk stratification, Transitional care, Discharge planning, Geriatric assessment, Polypharmacy, Logistic regression Introduction Background Thirty-day hospital readmissions in older adults often reflect suboptimal quality of care and fragmented care transitions. Multiple factors, including medical complexity, functional limitations, care dependency, and discharge organization, influence readmission risk in this population [1–3]. Inadequate coordination and unmet functional or social needs frequently lead to avoidable acute care use, reflecting missed opportunities for anticipatory discharge planning. These readmissions affect up to 20% of older adult discharges and are associated with increased morbidity, mortality, and healthcare costs [1, 2]. Numerous models have been developed to predict readmission risk, including the LACE index [4] developed in Canada, and the HOSPITAL score [5], derived in the United States (US). However, their performance in older adults is limited, with reported C-statistics often below 0.65 and as low as 0.58 in some geriatric subgroups [6–8]. Most were derived from general adult cohorts and omit geriatric-specific predictors, such as delirium risk, care dependency, and post-discharge support, despite their known relevance [3, 9–11]. For example, physical function, a key risk factor for post-discharge complications, is rarely included [3]. Machine learning (ML)-based models for heart failure similarly faced challenges with clinical interpretability, limiting their adoption [11]. Moreover, most models were developed and validated in North America, which differs substantially from the Swiss and broader European healthcare context. Switzerland’s acute care is decentralised, with multi-site hospital networks and no penalties for readmissions, unlike the US. Discharge coordination spans fragmented community providers, and post-acute rehabilitation and transitional care pathways are common. These features affect readmission drivers and model transferability, motivating a context-sensitive, locally validated approach. To address these gaps, we developed and internally validated a prediction model based on multivariable logistic regression. By integrating medical, organisational, geriatric, and care-dependency variables, the model aims to improve early identification of older adults at risk for unplanned 30-day readmission and to inform future transitional care interventions. Importantly, because readmissions and post-discharge mortality are competing outcomes, we also examined mortality risk separately to contextualise model performance. This prognostic study, part of a broader Medical Research Council (MRC) framework project, provides a quantitative foundation to inform co-designed hospital-to-home interventions [12]. Objectives The primary objective was to develop and internally validate a multivariable logistic regression-based model to predict unplanned hospital readmission within 30 days of discharge among patients aged ≥ 65 years. Secondary objectives were (1) modelling time to first readmission, and (2) predicting 30-day post-discharge mortality. Methods Study design and source of data We conducted a retrospective cohort study using routinely collected EHR data from the Ente Ospedaliero Cantonale (EOC), a public hospital group, structured as a centrum-level hospital network with four acute-care regional hospitals, serving about 350,000 residents in the Canton of Ticino, southern Switzerland [13]. The dataset included all hospitalisations of patients aged ≥ 65 years who were admitted and discharged between 1 January and 31 December 2024. Follow-up for 30-day readmissions and post-discharge mortality extended through 31 January 2025. The study protocol was prospectively registered on the Open Science Framework (OSF: https://osf.io/vkte2/) .The study was reported in accordance with the TRIPOD + AI guidelines [14], which provides harmonised standards for reporting prediction model studies using regression or machine-learning methods. Ethical approval for the overall project was granted by the Ethics Committee of the Canton Ticino (Protocol 2025-00025, Rif. CE 475). The requirement for informed consent was waived because of the retrospective use of de-identified routine data. Participants Eligible participants were all patients aged ≥ 65 years who were discharged alive to home, with or without home care. We excluded patients with planned readmissions (e.g. scheduled procedures), those admitted to oncology or palliative care units, and patients who died during the index hospitalisation. Admissions for post-acute rehabilitation and short observation stays were also excluded, as these do not represent acute unplanned readmissions. These criteria were applied to focus the analysis on unplanned readmissions that are more likely to reflect gaps in discharge planning or care quality [7]. Outcome measures The primary outcome was unplanned hospital readmission within 30 days of discharge, identified from EHR admission records. U nplanned 30-day readmission was defined as any acute, unscheduled inpatient admission to the same hospital network (EOC) within 30 days of discharge. This timeframe is a commonly used quality indicator in Organisation for Economic Co-operation and Development (OECD) countries [15, 16]. Planned admissions, scheduled procedures, post-acute rehabilitation, and palliative care admissions were excluded. Emergency department (ED) visits not resulting in admission and short observation stays were not counted as readmissions. Readmissions to hospitals outside the EOC network were not captured. Secondary outcomes were (1) time to first readmission (in days) and (2) 30-day post-discharge mortality. Mortality was analysed separately to account for its potential role as a competing event for readmission. Candidate predictors Predictors were selected a priori based on clinical relevance, prior literature, and routine availability in standardised EHR discharge forms completed by nurses, physicians, and administrative staff (Table 1). Candidate predictors included demographic, clinical, geriatric, hospital-related, and post-discharge factors. Geriatric measures comprised the Care Intensity Score (IAS), fall risk, and delirium risk, which were routinely assessed by nursing staff. The IAS is a 5-domain composite measure derived from the ePA-AC© [17] nursing assessment system and was used as a proxy for care dependency during hospitalisation, with scores ranging from 0 (low needs) to 14 (high needs). Higher scores indicate greater care complexity. Reference categories for categorical variables are summarised in Appendix 1. Table 1Candidate predictors and rationaleDomainPredictorRationaleDemographics aAge; Sex b; Marital status; Living situationDemographic factors shape vulnerability and support needs. Older age is associated with frailty and complications [18]. Sex may influence recovery trajectories, for example, men have been found to delay care-seeking, while women more often assume caregiving responsibilities, which may affect their own health outcomes [19–21]. Being partnered supports continuity of care [22], while living alone may hinder adherence and follow-up [23].Clinical statusPrimary diagnosis (ICD-10); comorbidity; DRG category; Intensive Care Unit (ICU) admission; polypharmacyClinical complexity and illness severity affect readmission risk. Chronic diseases (e.g., HF, COPD), ICU stay, and polypharmacy (≥ 5 medications/day) are known predictors of adverse events [24, 25]. ICD-10 and DRG help standardise disease burden and resource use.Geriatric assessmentsDelirium risk; fall risk; nutrition risk; care intensity score (IAS)Geriatric risks reflect patient vulnerability. Delirium, fall, and malnutrition are associated with frailty and readmission. [26]. The IAS score quantifies care complexity across five domains and informs discharge needs (Appendix 2)Hospital stayLength of stay; Admission type; WardHospital-related variables reflect care intensity and context. Longer stays often signal greater acuity [27]. Admission type and ward help control for institutional and specialty-related variation in care.Post-dischargeDischarge destination; Home care support at dischargeDischarge to home often requires follow-up; poor planning increases readmission risk [7]. Homecare services help reduce this risk by providing needed support [28].Outcomes30-day readmission; time to first readmission; 30-day post-discharge mortalityReadmission within 30 days is a widely used indicator [29]. Time to readmissions enables survival analysis. Mortality after discharge benchmarks continuity of care [29].Each predictor was available at discharge and had a plausible or established association with readmission risk in older adults.Abbreviations: ICD-10 International Classification of Diseases, Tenth Revision, ICU Intensive Care Unit, HF Heart Failure, COPD Chronic Obstructive Pulmonary Disease, DRG Diagnosis-Related GroupaRace and ethnicity were not collected, as these data are not routinely recorded in the Swiss hospital system and were unavailable for this analysisbSex refers to biological classification as recorded in the electronic health record; gender identity was not assessed [30] Candidate predictors and rationale Each predictor was available at discharge and had a plausible or established association with readmission risk in older adults. Abbreviations: ICD-10 International Classification of Diseases, Tenth Revision, ICU Intensive Care Unit, HF Heart Failure, COPD Chronic Obstructive Pulmonary Disease, DRG Diagnosis-Related Group aRace and ethnicity were not collected, as these data are not routinely recorded in the Swiss hospital system and were unavailable for this analysis bSex refers to biological classification as recorded in the electronic health record; gender identity was not assessed [30] Sample size The eligible sample included 15,635 hospitalizations. With a 30-day readmission rate of 19.6%, the sample supports inclusion of approximately 15–20 predictors, consistent with the recommended 10–20 events-per-variable (EPV) rule for prognostic model development [31, 32]. Although a formal power calculation was not performed, this EPV-based justification is widely accepted for sample size determination in prediction model research and indicates that the study had an adequate sample size to meet its objectives [31, 32]. Missing data Of the 15,635 eligible hospital discharges, 9,429 (60.3%) were included in the final analysis after excluding records with missing values for retained predictors. Nutrition risk had 61% missingness and failed MCAR (Little’s test p < 0.001), likely reflecting workflow variability and was excluded from the primary model. This exclusion reduced the analytic sample and may have introduced selection bias. Most retained predictors had low missingness (< 5%) in the analytic dataset; in, although in the full cohort, some variables (e.g., marital status, home care) exceeded 5%. We therefore applied complete-case analysis, while acknowledging its limitations, for transparency [33]. Statistical analysis All analyses were conducted in R version 4.5.0 (R Foundation for Statistical Computing, Vienna, Austria) (main packages: glmnet, caret, randomForest, xgboost, survival, pROC, car, ResourceSelection, rmda). Continuous predictors were entered on their original scale, and categorical variables were dummy coded with reference categories reported in Appendix 1. Primary model We developed a multivariable logistic regression model using backward stepwise guided by Akaike Information Criterion (AIC) selection, prioritising parsimony and interpretability [31]. To address potential instability, we ran a robustness analysis with LASSO (nested cross-validation, 200 bootstrap resamples) and addressed multicollinearity via variance inflation factors (VIF > 5 = problematic). Comparators As a baseline comparator, we implemented a simplified LACE index (LOS, comorbidity count as a proxy for the Charlson index, and admission acuity). The HOSPITAL score was not calculated due to missing components (see Appendix 3). Non-linear comparators included Random Forest (RF) and Extreme Gradient Boosting (XGBoost), with hyperparameters tuned using repeated k-fold cross-validation. Full tuning grids are reported in Appendix 4. LASSO, ridge regression and k-nearest neighbours (kNN) were run as sensitivity analyses. Performance assessment Model performance was evaluated by AUC (95% CI, bootstrap), calibration (slope, intercept, Brier score, Hosmer–Lemeshow test), and DCA, sensitivity, specificity, positive (PPV) and negative predictive values (NPV), confusion matrices, and precision–recall (PR) curves. Differences in AUC between models were tested using DeLong’s test with Holm correction. Internal validation was performed using 200 bootstrap resamples [32]. Secondary analyses and competing risks Time to first readmission was analysed using ordinary least squares (OLS) regression. Because 30-day post-discharge mortality was rare, competing-risk bias was expected to be limited. Nevertheless, we performed a Fine-Gray subdistribution hazards model treating death as a competing event. Results Participants characteristics Of the 15,635 patients aged ≥ 65, 12,814 (82.0%) met the inclusion criteria. After applying all exclusion criteria, including planned readmissions (n = 511, 3.3%), oncology or palliative care admissions (n = 1,210, 7.7%), in-hospital deaths (n = 1,100, 7.0%), and restricting the dataset to records with complete predictor information 9,429 patients (60.3%) remained for complete-case analysis. Mean age was 78.0 years (SD 7.6), 45.5% were female. Overall, 1,845 patients (19.6%) were readmitted within 30days. The cohort was clinically complex (median 6 comorbidities), and more than four out of five patients exposed to polypharmacy. Table 2 shows characteristics by readmission, with diagnostic details in Appendix 2 and mortality in Appendix 5. Readmitted patients were more likely to be male and to have greater comorbidity, polypharmacy, delirium risk, care intensity, and home care needs. Age and LOS did not differ between groups. The baseline differences were consistent with predictors retained in the regression model reinforcing their clinical plausibility. Table 2Baseline characteristics by 30-day readmission statusDomainPredictorNo readmission(n = 7 584)Readmission(n = 1 845)p-valueDemographicsAge, mean ± SD, y78.0 ± 7.678.0 ± 7.60.99Female sex, n (%)3 579 (47.2)709 (38.4)< 0.001Partnered, n (%)5 131 (67.6)1 208 (65.5)0.08Widowed, n (%)1 345 (17.7)380 (20.6)0.01Separated, n (%)1 108 (14.6)257 (13.9)0.42Living with others, n (%)5 124 (67.6)1 253 (67.9)0.82Living alone, n (%)2 460 (32.4)592 (32.1)0.82Clinical statusComorbidity count,median (IQR)6 (3–10)6 (3–10)< 0.001aICU admission, n (%)546 (7.2)164 (8.9)0.02Polypharmacy ≥ 5 medications,n (%)6 075 (80.1)1 716 (93.0)< 0.001Geriatric assessmentsDelirium risk present, n (%)6 392 (84.3)1 625 (88.1)< 0.001IAS (care intensity), n (%)< 0.001– Low4 270 (56.3)849 (46.0)-– Medium2 979 (39.3)891 (48.3)-– High335 (4.4)105 (5.7)-Fall risk, n (%)0.001– Low1 998 (26.4)439 (23.8)-– Medium3 719 (49)840 (45.5)-– High1 867 (24.6)566 (30.7)-Hospital stayLength of stay,median (IQR), days5 (3–8)5 (3–8)0.19Post-dischargeHome care support at discharge, n (%)1 853 (24.4)728 (39.5)< 0.001Baseline characteristics by 30-day readmission status. Values are presented as mean ± SD, median (IQR), or n (%). P-values were calculated using X2(test (or Fisher’s exact test, where appropriate) for categorical variables and the Wilcoxon rank-sum test for continuous variables. ICD-10 primary diagnosis chapters and DRG categories are detailed in Appendix 6. Percentages are based on available data; totals may not sum because of missing valuesAbbreviations: IQR Interquartile Range, SD Standard Deviation, ICD-10 International Classification of Diseases, Tenth Revision, ICU Intensive Care Unit, IAS Care Intensity Score Baseline characteristics by 30-day readmission status Comorbidity count, median (IQR) Polypharmacy ≥ 5 medications, n (%) Length of stay, median (IQR), days Baseline characteristics by 30-day readmission status. Values are presented as mean ± SD, median (IQR), or n (%). P-values were calculated using X2(test (or Fisher’s exact test, where appropriate) for categorical variables and the Wilcoxon rank-sum test for continuous variables. ICD-10 primary diagnosis chapters and DRG categories are detailed in Appendix 6. Percentages are based on available data; totals may not sum because of missing values Abbreviations: IQR Interquartile Range, SD Standard Deviation, ICD-10 International Classification of Diseases, Tenth Revision, ICU Intensive Care Unit, IAS Care Intensity Score Exploratory findings Comorbidity correlated with LOS (r = 0.57). Readmission was weakly correlated with comorbidity (r = 0.23) and ED visits (r = 0.30). Time to readmission was inversely related to readmission status (r = − 0.50). Age was weakly correlated with comorbidity (r = 0.21). Details are reported in Appendix 7. Model development Backward AIC selection retained 17 predictors (full coefficients in Appendix 8). Eight predictors had 95% CI that excluded the null: comorbidity count, male sex, polypharmacy, home care at discharge, lower fall risk, medium care intensity, younger age, and shorter LOS. Several other variables (e.g. ICU admission, marital status, hospital site) were retained by the selection procedure , despite wider CIs. These variables improved overall model fit but were not interpreted as clinically meaningful individual predictors (Appendix 8). The events-per-variable (EPV) ratio was > 100, reducing concerns about overfitting. Variance inflation factors for all predictors were < 5, indicating no relevant multicollinearity. A robustness analysis using LASSO regression identified a largely similar set of predictors (Appendix 8), supporting the stability of the model (Table 3). Table 3Multivariable logistic regression for 30-day readmissionPredictorOR95% CIp-valueComorbidity count (per diagnosis)1.121.10–1.14< 0.001Male sex1.421.27–1.60< 0.001Polypharmacy (≥ 5 medications)1.831.51–2.21< 0.001Home care at discharge1.611.43–1.81< 0.001Fall risk - Low vs. High1.411.23–1.61< 0.001Care intensity score (IAS) - Moderate vs. High1.301.01–1.670.047Age (per year)0.990.97–1.000.007Length of stay (per day)0.980.97–0.990.004Model metrics Intercept–2.58—— AIC8 512.5——Multivariable LR for 30-day readmission. Odds ratios (OR) CI, and p-values are shown for retained predictors with statistically significant associations in the final model. Reference categories: High fall risk, High care intensity (IAS), Female sex, < 5 medications, and no home care. Continuous predictors were modelled per unit increase. Model metrics (intercept and AIC) are listed. Full model including non-significant predictors is provided in Appendix 8Abbreviations: OR Odds ratio, CI Confidence Interval, IAS Care Intensity Score, AIC Akaike Information Criterion Multivariable logistic regression for 30-day readmission Multivariable LR for 30-day readmission. Odds ratios (OR) CI, and p-values are shown for retained predictors with statistically significant associations in the final model. Reference categories: High fall risk, High care intensity (IAS), Female sex, < 5 medications, and no home care. Continuous predictors were modelled per unit increase. Model metrics (intercept and AIC) are listed. Full model including non-significant predictors is provided in Appendix 8 Abbreviations: OR Odds ratio, CI Confidence Interval, IAS Care Intensity Score, AIC Akaike Information Criterion Model specification The probability (p) of 30-day readmission for an individual patient can be estimated using the logistic regression formula: where p is the probability of 30-day readmission, β0 is the intercept (–2.58), βi are the regression coefficients, and Xi are the values of patient-specific predictors. To illustrate how this formula can be applied, we provide an example calculation for a hypothetical patient (Table 4.) Table 4Worked exampleA 78-year-old man with 8 comorbidities and polypharmacy (≥5 medications), discharged with home care, moderate care intensity, low fall risk, and a 7-day hospital stay would have the following predicted risk of 30-day readmission:logit(p) =−2.58 + (−0.01×78) + (0.11×8) + 0.35 +0.60 + 0.48 + 0.34 + 0.13 + (−0.02×7) = −0.72 Thus, the predicted probability of 30-day readmission for this patient is 33%. Worked example A 78-year-old man with 8 comorbidities and polypharmacy (≥5 medications), discharged with home care, moderate care intensity, low fall risk, and a 7-day hospital stay would have the following predicted risk of 30-day readmission: logit(p) =−2.58 + (−0.01×78) + (0.11×8) + 0.35 +0.60 + 0.48 + 0.34 + 0.13 + (−0.02×7) = −0.72 Thus, the predicted probability of 30-day readmission for this patient is 33%. Model performance The logistic regression model achieved an AUC of 0.716 (95% CI 0.706–0.726). Internal validation using 200 bootstrap replications indicated minimal optimism (bias = 0.0015; standard error (SE) = 0.0062). Calibration of the original model showed systematic miscalibration (intercept − 2.60, slope 5.45, Brier 0.172, Hosmer-Lemeshow X²= 27.2, p < 0.001), indicating overestimation of risk at lower predicted probabilities. This pattern reflects substantial overestimation of absolute risk across most of the distribution, despite acceptable discrimination. Figure 1 shows calibration before recalibration. In addition, the Precision-Recall (PR) curve (Fig. 2) highlights limited sensitivity for identifying readmitted patients, with a PR-AUC of 0.364, consistent with the severe class imbalance. Calibration plot of the logistic regression model for 30-day readmission. The x-axis shows predicted risk by decile, the y-axis shows observed event rates. The dashed diagonal indicates perfect calibration Precision-recall curve for the logistic regression model predicting 30-day readmission. The x-axis shows recall (sensitivity), and the y-axis shows precision (positive predictive value) Competing risks Fine-Gray competing-risks regression, treating post-discharge mortality as a competing event, confirmed the main findings. Higher comorbidity was associated with an increased risk of 30-day readmission (sHR 1.11, 95% CI 1.10–1.12, p < 0.001), LOS was protective (sHR 0.98, 95% CI 0.97–0.99, p < 0.001). Age and ICU were not significant (Appendix 9). Comparators The logistic regression model achieved a modest but statistically significant improvement in discrimination compared with the LACE index (AUC 0.716 vs. 0.707, 95% CI 0.694–0.720; DeLong p < 0.01, Holm-adjusted p < 0.01). Among machine learning methods, RF (AUC 0.710, 95% CI 0.682–0.739) and XGBoost (AUC 0.707, 95% CI 0.697–0.752) achieved discrimination similar to logistic regression. Ridge regression performed comparably (AUC 0.708, 95% CI 0.681–0.733), whereas LASSO (AUC 0.685, 95% CI 0.661–0.708) and kNN (AUC 0.696, 95% CI 0.670–0.721) were significantly outperformed by logistic regression (Holm-adjusted p = 0.01 and 0.04, respectively). Detailed pairwise comparisons are provided in Appendix 10. Classification performance At a 10% threshold, logistic regression model achieved sensitivity 7.5%, specificity 98.1%, PPV 49.5%, NPV 81.4%. The F1-Score was 0.11, highlighting poor balance between precision and recall. PR analysis also showed only modest precision at higher levels of sensitivity, reflecting outcome imbalance. Confusion matrix and detailed metrics are provided in Appendix 11. Decision curve analysis DCA showed net benefit between thresholds of 5% and 20%, outperforming “treat all” and “treat none”. At thresholds above 30% net benefit declined, indicating reduced utility for restrictive intervention targeting (Figure in Appendix 12). Risk estimates The model generates continuous risk estimates. No fixed risk categories were predefined, in line with recommendations to allow local adaptation [14]. Secondary outcomes Among 3,441 patients, older age was associated with a longer time to readmission (β = 0.71, p = 0.0006). In contrast, male sex, a greater number of comorbidities, discharge with home care services, and low fall risk were associated with a shorter time to readmission (Appendix 12). Other predictors were not statistically significant (Appendix 13). Among 9,429 discharges, 80 patients (0.85%) died within 30 days. Patients who died were older and had higher comorbidity burden, longer hospital stays, and more frequent polypharmacy than survivors (Appendix 5). Mortality was also more common among patients discharged with home care. The mortality prediction model showed good discrimination (AUC 0.790, 95% CI 0.743–0.838), but poor calibration (slope 10.88, intercept − 4.88) reflecting the very low number of events and severe class imbalance (Appendix 14a–b). Discussion We developed and internally validated a prediction model based on LR for 30-day unplanned readmission in older adults using routinely collected discharge data from nurses, physicians, and administrative staff. The model showed acceptable discrimination and retained clinically interpretable predictors spanning medical complexity, geriatric vulnerability, and transitional care needs. Calibration was poor with overestimating risk at lower predicted probabilities, limiting the use of absolute risk estimates without local recalibration, though relative risk ranking remained reliable. Fine–Gray competing-risks analysis supported the consistency of the main findings: comorbidity increased risk, LOS was protective, and age and ICU admission were not significant. These results suggest that mortality as a competing event does not substantially alter the main determinants of readmission in this cohort. Core predictors, such as comorbidity burden, polypharmacy, and home care status, are consistent with prior evidence [4, 5, 7]. Nurse-reported measures (care intensity, fall risk) provided geriatric-specific insight that is often absent from generic models [4, 6, 7]. Because these indicators are routinely collected, the model may be feasible for integration into future discharge workflows, pending local recalibration and external validation, to inform risk-stratified transitional care. Nutrition risk, a well-known geriatric predictor, could not be modelleddue to non-random missingness. Improving systematic assessment is a priority for future work. Unexpectedly, patients classified as having low fall risk or moderate care intensity had higher readmission odds than those classified as high risk. This may reflect an “inverse care” effect, whereby high-risk patients receive targeted interventions (e.g., rehabilitation, follow-up), whereas apparently lower-risk patients receive less structured support. This finding highlights the need to look beyond categorical labels and to consider unmeasured vulnerabilities, such as health literacy and caregiver availability [34, 35]. Compared with LACE and HOSPITAL, our model incorporates geriatric assessments and reflects the decentralised Swiss context. However, calibration remained imperfect, and sensitivity was uniformly low across models. This suggests that performance may be limited more by the available data than by modelling strategy [36, 37]. Algorithmic complexity did not compensate for the absence of data on functional recovery, social support, or outpatient use. DCA showed net clinical benefit across threshold probabilities between 5% and 20%, suggesting potential care programmes. For example, patients at very high predicted risk could be prioritised for bundled interventions (home nursing visits, medication review, community linkage), while those at intermediate risk might receive structured follow-up calls [38, 39]. This probability-based approach is consistent with best practices in transitional care [37, 38]. Secondary analyses added limited clinical insight: both time to readmission and 30-day mortality were only weakly explained by discharge-level predictors. For 30-day mortality, the model showed acceptable discrimination but substantial miscalibration because of the very low number of events and severe class imbalance. These results should therefore be interpreted as descriptive and hypothesis-generating only. Overall, the model provides transparent, context-sensitive predictions of readmission in older adults. Although absolute risks require recalibration and sensitivity is low, relative risk ranking may still support stratified transitional care and future integration into discharge summaries, in line with learning health system principles. Strengths and limitations Strengths include the use of a large, real-world dataset from a multi-site hospital network, clinically interpretable predictors, and the incorporation of geriatric variables such as the IAS and delirium risk. The inclusion of nurse-reported measures supports feasibility of use in routine practice. Reporting adhered to TRIPOD + AI, and the study was guided by the MRC framework and a preregistered protocol. A major limitation is the high level of missingness for nutrition risk (61%) and the consequent exclusion of 40% of patients, which may have introduced selection bias. Excluded patients may have differed systematically in frailty and readmission risk, potentially biasing both calibration and generalisability. Other limitations include the single-hospital network setting, which limits generalisability; the locally defined IAS score; and the absence of social determinants (income, education, access to primary care), which may left residual confounding [40, 41]. Readmissions to hospitals outside our network could not be captured, leading to potential under-ascertainment of events. The model also showed poor calibration, requiring local recalibration before clinical use. Machine learning comparators did not outperform logistic regression, underscoring the value of interpretable models, although some jointly predictive features may have been omitted through backward elimination [42]. We chose stepwise AIC for parsimony and interpretability and confirmed robustness using LASSO, which yielded a similar predictor set. Finally, routine EHR data may be subject to misclassification, although documentation was standardised. Implications Clinically, this model may help inform risk-stratified transitional care for older adults by providing continuous, interpretable risk estimates. Even with low sensitivity, DCA suggested potential value for relative risk stratification, enabling flexible, threshold-based planning [36, 38]. With further recalibration and evaluation, integration into discharge workflows and coordination platforms could be explored [43]. In future, such models might also be adapted dynamically by updating risk estimates with post-discharge information (e.g., symptom changes, missed appointments), in line with learning health system approaches [44, 45]. For research, the model contributes to the development of complex intervention under the MRC framework [12]. It may help identify patient subgroups who could benefit from enhanced transitional care and support the co-design of multi-component interventions. Future work should prioritise external validation, improved capture of geriatric and social determinants, and a prospective impact evaluation. Comparisons with established scores such as HOSPITAL, alongside strategies to improve calibration, will be essential to determine the model's added value. In larger cohorts, broader application of competing-risk methods may further refine understanding of readmission dynamics. Conclusions We developed and internally validated a pragmatic, interpretable model for predicting 30-day readmission in older adults. The model identified geriatric and care-related predictors, although calibration and sensitivity were suboptimal. With further refinement, recalibration, and external validation, the model may help inform resource planning and design of interventions to reduce avoidable readmissions. Supplementary Information Additional file 1. Contains supplementary appendices (1-14), including details on the Care Intensity Score (IAS), diagnostic coding, model development and validation results, decision curve analysis, sensitivity analyses, and secondary outcome results Additional file 2. TRIPOD+AI checklist Acknowledgements We thank Dr. Kay Stankov for his expertise in complex statistical analyses and his support in interpreting the results. We also thank Michele Villa for assistance with exploratory data analysis, Dante Cheda for valuable insights into the structure and use of EHR data. We also thank Amos Colombo for extracting and preparing the dataset. We are grateful to the Ente Ospedaliero Cantonale (EOC) for institutional support, and to University of Applied Sciences and Arts of Southern Switzerland (SUPSI) for providing time that enabled the first author to contribute to this work as part of her PhD. Patient and public involvement No involvement Abbreviations Akaike Information Criterion Area Under the Receiver Operating Characteristics Curve Confidence Interval Decision Curve Analysis Diagnosis-Related Group Emergency Department Electronic Health Record Ente Ospedaliero Cantonale elektronisches Patientendossier (electronic patient record) - Acute Care Events per Variable Hosmer-Lemeshow Hazard Ratio Care Intensity Score International Classification of Diseases, Tenth Revision Intensive Care Unit Interquartile Range k-Nearest Neighbours Least Absolute Shrinkage and Selection Operator Medical Research Council Organisation for Economic Cooperation and Development Ordinary Least Squares Odds Ratio Open Science Framework Random Forest Precision Recall Receiver Operating Characteristic Standard Deviation Standard Error subdistribution Hazard Ratio Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis + Artificial Intelligence Variance Inflation Factor Extreme Gradient Boosting Authors’ contributions Concept and design: all authors; Acquisition, analysis, or interpretation of data: all authors; Statistical analysis (conceptualization and oversight): LMS, LB, SMGZ, SH; Drafting of the manuscript: LMS; Critical revision of the manuscript for important intellectual content: SMGZ, SH, LB; Supervision: SMGZ, SH. Funding No direct funding was received for this study. However, data extraction and technical support were provided by EOC, which also supported the project through in-kind contributions, including staff time, and access to institutional resources. Data availability In line with TRIPOD+AI, a dummy dataset (.csv) and annotated R Markdown (.rmd) script are openly available on OSF: https://osf.io/vkte2. These are provided in the archive “ReadmissionModel_FAIRpackage.zip”, which also includes licenses (MIT for code, CC-BY for data) and a README. Declarations Ethics approval and consent to participate Ethical approval for the overall project was granted by the Ethics Committee of the Canton Ticino (Protocol 2025-00025, Rif. CE 475). Informed consent was waived due to the retrospective use of de-identified routine data. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Additional file 1. Contains supplementary appendices (1-14), including details on the Care Intensity Score (IAS), diagnostic coding, model development and validation results, decision curve analysis, sensitivity analyses, and secondary outcome results Additional file 2. TRIPOD+AI checklist Data Availability Statement In line with TRIPOD+AI, a dummy dataset (.csv) and annotated R Markdown (.rmd) script are openly available on OSF: https://osf.io/vkte2. These are provided in the archive “ReadmissionModel_FAIRpackage.zip”, which also includes licenses (MIT for code, CC-BY for data) and a README.