Development and validation of a clinical prediction model and risk score for 30-day psychiatric readmission in the United Arab Emirates

Abstract

Background

Psychiatric readmission within 30 days of discharge is an important indicator of healthcare quality. Despite the global proliferation of prediction models, no validated tool exists for Middle Eastern or Gulf Cooperation Council populations, where distinctive demographic characteristics and service configurations may limit the transportability of models developed elsewhere.

Methods

This retrospective cohort study analysed 14,994 psychiatric inpatient admissions from 9,263 unique patients at a tertiary psychiatric facility in the United Arab Emirates (2018–2025), using electronic health record data. The cohort was randomly split into development (70%) and validation (30%) samples. Missing marital status data (28.2%) were addressed using multiple imputation with chained equations (20 imputed datasets). Nonlinearity of count-based predictors was addressed using log-transformations. LASSO logistic regression with ten-fold cross-validation was used for predictor selection and model fitting, with results pooled across imputed datasets. A simplified integer-based clinical risk score was derived from categorised predictors. Model performance was benchmarked against recalibrated LACE and READMIT-subset indices.

Results

The 30-day readmission rate was 10.6% (n = 1,583). LASSO regression identified prior psychiatric admissions, schizoaffective disorder, Charlson Comorbidity Index, first-generation antipsychotic prescription, long-acting injectable antipsychotic prescription (protective), emergency admission, emergency department visits, and antidepressant prescription as the principal predictors. Marital status was not consistently retained after multiple imputation with married as the reference category, likely reflecting the large proportion of missing data and the characteristics of the unrecorded group. The model demonstrated modest discrimination (AUC 0.649, 95% CI 0.623–0.678 in validation), significantly outperforming both the LACE index (AUC 0.557, p < 0.001) and READMIT-subset (AUC 0.604, p < 0.001). Temporal validation (2018–2023 training, 2024–2025 testing) yielded stable performance (AUC 0.676). A three-tier risk score stratified patients into low (6.7%), medium (10.0%), and high (18.0%) readmission probability groups.

Conclusion

This study presents the first psychiatric readmission prediction model from the Middle East, developed using penalised regression and multiple imputation. The locally derived model outperformed imported indices, and the clinical risk score offers a practical tool for identifying high-risk patients who may benefit from intensified discharge planning and transitional care.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12888-026-08165-z.

Background

Psychiatric readmission within 30 days of discharge is widely used as a health-system performance indicator as well as a marker of continuity and effectiveness of post-discharge care in mental health services [1]. Reported rates vary substantially across settings and case-mix. For example, a territory-wide cohort in Hong Kong reported a 7.09% rate of 28-day unplanned readmission following discharge from acute psychiatric units, while system-level studies in other jurisdictions report materially higher short-term readmission burdens [2–5]. In a systematic review and meta-analysis of 13 studies, the pooled 30-day unplanned psychiatric readmission rate was 16% (95% confidence interval 13–20%), with higher readmission propensity observed in schizophrenia-spectrum disorders and personality disorders, and elevated risk also reported for bipolar disorder [6].

The consequences of psychiatric readmission extend beyond hospital utilisation to substantial societal and caregiver burden. Rehospitalisation contributes materially to healthcare costs in severe mental illness, and analyses of schizophrenia and psychotic-disorder admissions have highlighted that early readmissions are an important cost driver in U.S. hospital systems [7]. Beyond economic impact, psychiatric hospitalisation and readmission disrupt recovery and continuity of care and are associated with significant strain on family caregivers [8]. In a systematic review and meta-analysis, the pooled prevalence of clinically significant caregiver burden among caregivers of individuals with mental illness was approximately one-third, with higher levels reported in hospital and psychosis-related samples [9]. Caregivers of people with psychosis may also experience fear, trauma, and helplessness, particularly where patient aggression or violence occurs, with measurable adverse effects on caregiver wellbeing [10]. These human and financial consequences underscore the need for accurate, locally calibrated readmission risk prediction to enable targeted post-discharge interventions.

Current prediction models and their limitations

Several readmission prediction tools have been developed, yet routine implementation in psychiatric inpatient services remains limited. The LACE index (Length of stay, Acuity of admission, Comorbidity, Emergency department visits) was derived in general hospital populations and explicitly excluded psychiatric admissions from its derivation framework [11, 12]. The READMIT index (Repeat admissions, Emergent admission, Age, Diagnoses, Medical morbidity, Intensity, Time in hospital) was developed specifically for psychiatric inpatient populations in Ontario, Canada, demonstrating moderate discrimination (c-statistic 0.63) in derivation and validation cohorts, though external validation has yielded mixed results [13, 14].

Machine learning approaches have demonstrated improved discrimination in some settings, though methodological limitations remain pervasive: a systematic review of 228 psychiatric prediction model studies found that 94.5% were at high risk of bias, with external validation and implementation evaluation uncommon [15]. Importantly, transportability across health systems is limited; cross-site validation studies demonstrate consistent performance degradation when models are applied to populations differing from derivation cohorts, underscoring the need for locally calibrated models [5].

The Middle East and North Africa (MENA) research gap

Psychiatric readmission research in the Gulf Cooperation Council (GCC) remains limited, and the available evidence is dominated by retrospective single-centre descriptive studies rather than prediction model development and validation. For example, a tertiary-centre study in Oman reported a 39% 1-year readmission rate and examined associated clinical and sociodemographic factors, while a Qatar study described the characteristics of patients with recurrent psychiatric readmissions in a general hospital setting and reported a 30-day readmission proportion within its sample [16, 17]. Across the region, schizophrenia-focused readmission work has also been reported, but estimates vary substantially by cohort and design, underscoring the need for locally derived, methodologically robust tools [18, 19]. To our knowledge, there remains a gap in externally validated, deployable psychiatric 30-day readmission prediction tools tailored to Gulf health systems, including the United Arab Emirates.

This gap is particularly important given the distinctive service context. In the UAE, psychiatric bed capacity has been reported at approximately 0.9 beds per 100,000 population, markedly lower than figures reported for the United Kingdom (23.9 per 100,000) [20]. Workforce capacity has also been constrained; recent national reporting indicates psychiatrist density in the range of ~ 4–5 per 100,000 in 2020–2022 [21, 22]. In such settings, where care pathways, follow-up infrastructure, and population structure (including large expatriate communities) differ from Western derivation cohorts, transportability of imported prediction tools may be limited, strengthening the case for locally calibrated risk stratification.

Importantly, the rationale for developing a region-specific model does not rest on the expectation that fundamentally different predictor domains will emerge; core predictors such as prior hospitalisation, illness severity, and comorbidity are well established across settings. Rather, the need arises from the possibility that the relative importance of these predictors, their coefficient magnitudes, and the optimal risk thresholds may differ in a population characterised by younger age, high expatriate proportion, and distinctive service capacity constraints, necessitating locally derived coefficient estimates and risk thresholds rather than direct importation of models developed in Western healthcare systems.

Study rationale and objectives

Given evidence that psychiatric readmission prediction models can lose performance when applied across healthcare settings and sites [5], the limited published literature from the United Arab Emirates reporting validated 30-day psychiatric readmission risk prediction, and the substantial clinical and resource consequences associated with unplanned readmissions, we aimed to develop and internally validate a locally derived prediction model for unplanned 30-day psychiatric readmission following inpatient discharge. We hypothesised that a model incorporating locally available demographic, diagnostic, and service-utilisation variables would outperform imported generic indices such as the Length of stay, Acuity, Comorbidity, Emergency department visits index (LACE) [11] and psychiatric-specific tools developed in different healthcare systems, such as the READMIT index [13]. Additionally, we sought to translate the final model into a practical point-based clinical risk score capable of stratifying patients into clinically meaningful risk categories to inform discharge planning and guide the intensity of post-discharge follow-up and targeted interventions.

Methods

Study design and setting

This retrospective cohort study analysed all psychiatric inpatient admissions at a Al Amal Psychiatric Hospital, Dubai, United Arab Emirates from January 1, 2018, to December 22, 2025. Data were extracted from the hospital’s electronic health record system, which captures all clinical encounters including diagnoses coded using the International Classification of Diseases, Tenth Revision, Australian Modification (ICD-10-AM), medication orders, encounter dates, and demographic information. The study followed the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines for prediction model development and validation.

Study population

The initial database contained 224,969 total encounters from 25,337 patients, of which 15,363 were inpatient psychiatric encounters. Exclusion criteria included: admissions outside the study period (n = 30), transfer admissions within 24 h of prior discharge (n = 4), missing discharge date (n = 156), and inadequate 30-day follow-up (n = 179). The final analytic cohort comprised 14,994 admissions from 9,263 unique patients (Fig. 1). All 14,994 included encounters had a minimum 30-day post-discharge observation window. The cohort was randomly split 70:30 into development (n = 10,495) and validation (n = 4,499) sets, stratified by the primary outcome.

Fig. 1.

Cohort Selection and Study Flow Diagram. Note. Flow diagram depicting the stepwise derivation of the study cohort from the institutional database. Of 224,969 total encounters, 15,363 were inpatient psychiatric admissions. Sequential exclusions were applied for encounters outside the study period (n = 30), inter-facility transfers (n = 4), missing discharge dates (n = 156), and insufficient 30-day follow-up (n = 179), yielding a final analytic cohort of 14,994 admissions representing 9,263 unique patients. Random allocation to development (70%) and validation (30%) sets achieved balanced readmission rates (10.6% in both sets)

Outcome definition

The primary outcome was 30-day psychiatric readmission, defined as any subsequent psychiatric inpatient admission within 30 days of the index discharge date. Readmissions within 24 h were classified as transfers and excluded. Secondary outcomes included 7-day, 90-day, and 365-day readmission.

Candidate predictors

A total of 33 candidate variables were evaluated across four domains: (1) demographics (age, gender, nationality, marital status); (2) admission characteristics (emergency vs. elective, weekend discharge, length of stay); (3) clinical variables (primary psychiatric diagnosis, Charlson Comorbidity Index, healthcare utilisation history including emergency department visits and prior admissions); and (4) discharge medications (first-generation antipsychotics [FGA], second-generation antipsychotics [SGA], mood stabilisers, lithium, antidepressants, benzodiazepines, anticholinergics, long-acting injectable antipsychotics [LAI], polypharmacy status). Discharge medication variables were included a priori based on clinical rationale: prescribing patterns at discharge may serve as markers of illness severity and treatment complexity, and specific formulations such as LAIs have been associated with reduced rehospitalisation in observational studies [23]. Their inclusion enabled a pre-specified assessment of whether pharmacotherapy data improves prediction beyond demographic and clinical variables.

Statistical analysis

Missing data handling

Marital status was unrecorded for 4,227 encounters (28.2%). The pattern of missingness was assessed by fitting a logistic regression model predicting missing marital status from all other candidate predictors. Missingness was significantly associated with several observed variables (pseudo R² = 0.075), including nationality, prior admissions, and admission year, supporting a Missing At Random (MAR) assumption (Supplementary Table S1–S3). Multiple imputation using chained equations was performed with 20 imputed datasets, employing Bayesian Ridge regression as the imputation model with all candidate predictors and the outcome variable included in the imputation model [24]. The imputed marital status variable was coded with married as the reference category, as the most common and clinically intuitive baseline. A complete-case sensitivity analysis excluding encounters with unrecorded marital status was also performed.

Functional form of continuous predictors

The linearity assumption for continuous predictors was assessed by comparing Akaike Information Criterion values across linear, quadratic, and logarithmic functional forms. Count-based utilisation variables (prior admissions, emergency department visits, outpatient visits, medication count) demonstrated evidence of nonlinearity (ΔAIC > 6 for five of six variables; Supplementary Table S4). These variables were log-transformed using log(x + 1) for modelling, as this is the standard approach for right-skewed count data in clinical prediction models and yielded comparable fit to quadratic terms with greater interpretability.

Model development

LASSO (Least Absolute Shrinkage and Selection Operator) logistic regression with L1 penalisation was used for simultaneous predictor selection and coefficient estimation, as recommended by contemporary guidelines for clinical prediction model development [25]. All 33 candidate predictors were entered directly into the LASSO model without prior univariable screening, as penalised regression methods handle variable selection internally through regularisation. The optimal regularisation parameter was determined by ten-fold cross-validation optimising the area under the receiver operating characteristic curve (AUC). The model was fitted separately on each of the 20 imputed datasets, and coefficients were pooled using Rubin’s rules. Variables retained in fewer than 50% of imputed datasets were considered not selected by the model. For comparison, Elastic Net (combined L1 and L2 penalisation) and Ridge (L2 penalisation) logistic regression models were also fitted (Supplementary Table S5). The LASSO regularisation path illustrating predictor entry as penalisation decreases is shown in Supplementary Fig. S1.

Model performance was evaluated in the validation set. Discrimination was assessed using the AUC with 1,000-iteration bootstrap 95% confidence intervals. Calibration was evaluated using calibration-in-the-large, calibration slope, Hosmer-Lemeshow goodness-of-fit test, and calibration plots (Supplementary Fig. S2). Overall performance was quantified using the Brier score. Clinical utility was assessed via decision curve analysis (Supplementary Fig. S3). Statistical significance was defined as two-sided p < 0.05. All analyses were performed using Python 3.x with scikit-learn, statsmodels, and scipy.

Benchmark comparisons

Model performance was compared against two established readmission prediction instruments: the LACE index and the READMIT index. The READMIT implementation excluded two unavailable components (threat to self/others, unplanned discharge), resulting in a READMIT-subset score. Both benchmarks were recalibrated to the local population using logistic regression in the development set. Formal AUC comparisons used bootstrap-based variance estimation (Supplementary Table S6). A structured predictor-level comparison between the local model and the READMIT index is presented in Table 1.

Table 1.

Structured Comparison of predictors: READMIT Index vs. Local Model

Domain	READMIT Index	Local Model	Difference
Prior utilisation	Repeat admissions (R)	Prior admissions lifetime (log) + 12 m (log)	Log-transformed; two time windows
Emergency	Threat to self/others	Emergency admission (binary)	READMIT uses danger criteria (unavailable in data)
Age	Age categories (A)	Age (continuous, linear)	Continuous vs. categorical
Diagnosis	Diagnosis group + personality + unplanned DC	Schizoaffective, personality, neurodevelopmental	Separates schizoaffective from other psychoses
Comorbidity	Charlson (M)	Charlson (continuous)	Continuous vs. categorical
Service use	OP visits + ED visits (I)	ED visits 6 m (log)	OP visits not selected by LASSO
LOS	Time in hospital (T)	Not selected	LOS not predictive after adjustment
Medications	Not included	FGA (+), LAI (−), antidepressant (+)	Novel medication markers

Note. READMIT = Repeat admissions, Emergent admission, Age, Diagnoses, Medical morbidity, Intensity, Time in hospital. The READMIT implementation excluded two unavailable components (threat to self/others, unplanned discharge). LOS, length of stay; OP, outpatient; ED, emergency department; FGA, first-generation antipsychotic; LAI, long-acting injectable antipsychotic; DC, discharge

Reclassification and incremental value

Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) quantified classification improvements (Supplementary Table S7). The incremental predictive value of discharge medications was assessed by comparing a clinical-only model (Model A) against the full model with medications (Model B) (Supplementary Table S8).

Clinical risk score development and validation

A simplified integer-based clinical risk score was derived to facilitate bedside application. Continuous predictors retained by LASSO were first categorised into clinically meaningful groups based on observed readmission rate gradients: prior admissions (0–1, 2–3, 4+), emergency department visits (0, 1, 2+), and Charlson Comorbidity Index (0, ≥ 1). An unpenalised logistic regression model was then fitted using these categorised predictors alongside the binary predictors selected by LASSO. Integer point values were assigned by dividing each regression coefficient by the smallest significant coefficient and rounding to the nearest integer. Risk tiers (low, medium, high) were defined using score tertiles in the development cohort. Long-acting injectable antipsychotic prescription received negative points, reflecting its protective association. The discriminative ability of the integer risk score was compared against the full logistic model, and internal validation was performed using 500 bootstrap iterations to estimate optimism-corrected AUC values (Supplementary Table S9).

Secondary and sensitivity analyses

Secondary analyses examined alternative time horizons (7-day, 90-day readmission), diagnostic subgroups (schizophrenia, bipolar disorder, substance use disorders, schizoaffective disorder, depression), and recurrent readmissions using negative binomial regression (Supplementary Table S10). A dedicated first-admission-only LASSO model was fitted using only variables available at index presentation (excluding prior utilisation history) to assess predictive capacity for patients without established readmission patterns (Supplementary Table S11).

Sensitivity analyses included: (1) temporal validation (development on 2018–2023, validation on 2024–2025); (2) exclusion of short stays (< 48 h); (3) restriction to the post-COVID period (2020–2025); (4) complete-case analysis excluding encounters with unrecorded marital status; and (5) the original AIC-based backward stepwise model for comparison with the LASSO approach (Supplementary Table S12). Competing risks from post-discharge mortality were addressed through a quantitative bounding analysis using published 30-day post-discharge psychiatric mortality rates (0.5–1.5%), as individual-level mortality data were not available in the electronic health record extract (Supplementary Table S13).

Results

Study Population and readmission rates

The study cohort derivation is shown in Fig. 1. The final cohort comprised 14,994 admissions from 9,263 unique patients spanning eight years. The overall 30-day readmission rate was 10.6% (1,583/14,994), with identical rates in development (10.6%) and validation (10.6%) sets. Secondary outcome rates were: 7-day 3.3%, 90-day 17.8%, and 365-day 29.5%.

Baseline characteristics

Table 2 presents baseline characteristics stratified by 30-day readmission status, with standardised mean differences (SMD) used to quantify imbalances between groups. The median age was 31 years (IQR 26–39), and 32.7% were female. Emergency admissions accounted for 81.0% of hospitalisations. The most common diagnoses were substance use disorders (33.5%), bipolar disorder (24.6%), schizophrenia (16.1%), and depression (9.1%).

Table 2.

Baseline Characteristics of the study cohort stratified by 30-day readmission status

Variable	No Readmission (n = 13,411)	Readmitted (n = 1,583)	SMD
Age, years	31.0 (26.0–39.0)	31.0 (26.0–40.0)	0.030
Length of stay, days	11.4 (5.6–21.9)	12.3 (5.8–23.0)	0.045
Charlson Comorbidity Index	0.0 (0.0–0.0)	0.0 (0.0–0.0)	0.196
ED visits, prior 6 months	0.0 (0.0–0.0)	0.0 (0.0–1.0)	0.248
ED visits, prior 12 months	0.0 (0.0–1.0)	0.0 (0.0–1.0)	0.275
Prior admissions, 12 months	1.0 (1.0–2.0)	1.0 (1.0–3.0)	0.613
Prior admissions, lifetime	1.0 (1.0–2.0)	2.0 (1.0–4.0)	0.645
Outpatient visits, prior 12 months	0.0 (0.0–2.0)	1.0 (0.0–5.0)	0.279
Medication count at discharge	6.0 (5.0–7.0)	6.0 (5.0–7.0)	0.207
Female sex	4,381 (32.7%)	524 (33.1%)	0.009
UAE national	12,027 (89.7%)	1,441 (91.0%)	0.046
Single	5,958 (44.4%)	844 (53.3%)	0.179
Married	3,342 (24.9%)	418 (26.4%)	0.034
Emergency admission	10,779 (80.4%)	1,370 (86.5%)	0.167
Weekend discharge	2,439 (18.2%)	297 (18.8%)	0.015
Schizophrenia	2,149 (16.0%)	269 (17.0%)	0.026
Schizoaffective disorder	921 (6.9%)	210 (13.3%)	0.214
Bipolar disorder	3,303 (24.6%)	391 (24.7%)	0.002
Depression	1,216 (9.1%)	155 (9.8%)	0.025
Anxiety disorder	395 (2.9%)	53 (3.3%)	0.023
Substance use disorder	4,512 (33.6%)	512 (32.3%)	−0.028
Personality disorder	658 (4.9%)	107 (6.8%)	0.079
Neurodevelopmental disorder	564 (4.2%)	121 (7.6%)	0.146
First-generation antipsychotic	9,727 (72.5%)	1,237 (78.1%)	0.130
Second-generation antipsychotic	11,853 (88.4%)	1,402 (88.6%)	0.006
Mood stabiliser	4,667 (34.8%)	674 (42.6%)	0.160
Lithium	409 (3.0%)	91 (5.7%)	0.132
Antidepressant	4,143 (30.9%)	560 (35.4%)	0.095
Benzodiazepine	10,043 (74.9%)	1,218 (76.9%)	0.048
LAI at discharge	1,756 (13.1%)	209 (13.2%)	0.003
Polypharmacy (≥ 3 classes)	12,931 (96.4%)	1,533 (96.8%)	0.023
7-day outpatient follow-up	3,037 (22.6%)	365 (23.1%)	0.010

Note. Values are presented as median (interquartile range) for continuous variables and n (%) for categorical variables. Imbalances are quantified using standardised mean differences (SMD); an absolute SMD > 0.1 indicates a meaningful difference between groups. IQR, interquartile range; ED, emergency department; LAI, long-acting injectable antipsychotic

The largest imbalances between readmitted and non-readmitted patients were observed for prior psychiatric admissions (lifetime: SMD 0.645; 12-month: SMD 0.613), outpatient visits (SMD 0.279), emergency department visits (12-month: SMD 0.275; 6-month: SMD 0.248), medication count (SMD 0.207), schizoaffective disorder (SMD 0.214), and Charlson Comorbidity Index (SMD 0.196). Readmitted patients were also more likely to have emergency admissions (SMD 0.167), single marital status (SMD 0.179), and first-generation antipsychotic prescriptions (SMD 0.130). Age, sex, bipolar disorder, substance use disorder, and weekend discharge showed negligible imbalances (SMD < 0.05).

Marital status was unrecorded for 4,227 encounters (28.2%). Characterisation of the unrecorded group revealed systematic differences: compared with patients with recorded marital status, those with missing data were less likely to be UAE nationals (84.4% vs. 91.9%), had fewer prior admissions, and had a lower observed 30-day readmission rate (7.0% vs. 12.0%). Missingness was significantly predicted by non-UAE nationality, fewer prior admissions, and later admission year (Supplementary Tables S1 and S14), consistent with a Missing At Random mechanism and potentially reflecting documentation practices for patients with shorter system engagement. Multiple imputation was used to address this missingness (see Methods).

Multivariable model

LASSO logistic regression, applied across 20 multiply imputed datasets with log-transformed count variables, retained 24 of 33 candidate predictors with nonzero coefficients in at least 50% of imputed datasets (Table 3). The strongest predictors, ranked by pooled standardised coefficient magnitude, were: prior psychiatric admissions (lifetime, log-transformed; pooled coefficient 0.314), first-generation antipsychotic prescription at discharge (0.138), prior admissions in the preceding 12 months (log-transformed; 0.128), schizoaffective disorder (0.117), Charlson Comorbidity Index (0.095), and long-acting injectable antipsychotic prescription (− 0.094, protective). Additional predictors with meaningful coefficients included antidepressant prescription (0.059), emergency department visits in the prior 6 months (log-transformed; 0.058), emergency admission (0.053), and younger age (− 0.044).

Table 3.

LASSO logistic regression model for 30-day psychiatric readmission: pooled coefficients across 20 multiply imputed datasets

Variable	Pooled Coefficient	% Nonzero	Direction
Prior admissions, lifetime (log)	+ 0.3143	100%	Positive
First-generation antipsychotic	+ 0.1376	100%	Positive
Prior admissions, 12 months (log)	+ 0.1281	100%	Positive
Schizoaffective disorder	+ 0.1166	100%	Positive
Charlson Comorbidity Index	+ 0.0950	100%	Positive
LAI at discharge	−0.0937	100%	Negative (protective)
Antidepressant	+ 0.0586	100%	Positive
ED visits, 6 months (log)	+ 0.0581	100%	Positive
Marital other vs. married	−0.0558	95%	Negative
Emergency admission	+ 0.0531	100%	Positive
Age	−0.0439	100%	Negative (younger = higher risk)
Personality disorder	+ 0.0284	100%	Positive
Neurodevelopmental disorder	+ 0.0215	100%	Positive
Mood stabiliser	+ 0.0193	100%	Positive
Weekend discharge	−0.0159	100%	Negative
Lithium	+ 0.0128	100%	Positive
Anxiety disorder	+ 0.0106	100%	Positive
ED visits, 12 months (log)	+ 0.0081	100%	Positive
Depression	+ 0.0079	100%	Positive
Second-generation antipsychotic	−0.0079	90%	Negative
Single vs. married	+ 0.0054	50%	Positive (not retained)
Schizophrenia	+ 0.0048	85%	Positive
UAE national	−0.0023	85%	Negative
Any antipsychotic	+ 0.0014	80%	Positive

Note. Coefficients are on the standardised scale. Variables ranked by absolute pooled coefficient magnitude. “% Nonzero” indicates the proportion of imputed datasets in which the variable received a nonzero coefficient. Variables retained in ≥ 50% of imputations were considered selected. Married status served as the reference category for marital status following multiple imputation. Count-based variables were log-transformed using log(x + 1)

Notably, single marital status was retained in only 50% of imputed datasets with a near-zero pooled coefficient (0.005), indicating that the strong apparent association observed in preliminary analyses (aOR 1.45 with unrecorded marital status as reference) was substantially an artefact of the missing-data reference category. After multiple imputation with married as the reference, marital status did not meaningfully contribute to readmission prediction. The divorced/widowed/separated category showed a modest protective association (coefficient − 0.056, retained in 95% of imputations).

Nine candidate predictors were not selected by LASSO: female sex, substance use disorder (as primary diagnosis), bipolar disorder, length of stay, outpatient visits, 7-day follow-up, benzodiazepine prescription, polypharmacy, and overall antipsychotic prescription Univariable associations for all 33 candidate predictors are presented in Supplementary Table S15. Comparison of LASSO, Elastic Net, Ridge, and the original AIC-based model demonstrated virtually identical discrimination (all AUCs 0.649–0.652), indicating that the original model was not substantially overfit and that the revised penalised approach confirms the robustness of the initial findings (Supplementary Table S5).

Model performance and benchmark comparisons

In the validation set (n = 4,499), the revised LASSO model achieved an AUC of 0.649 (95% CI 0.623–0.678) with a Brier score of 0.091 (Table 4). The model significantly outperformed both benchmarks: recalibrated LACE (AUC 0.557; ΔAUC = 0.092, z = 5.69, p < 0.001) and recalibrated READMIT-subset (AUC 0.604; ΔAUC = 0.045, z = 3.47, p < 0.001). Figure 2 displays the comparative ROC curves. Compared with LACE, the local model demonstrated significant net reclassification improvement (continuous NRI = 0.400, categorical NRI at 10% threshold = 0.137) and integrated discrimination improvement (IDI = 0.035). Against the READMIT-subset, continuous NRI was 0.290 and IDI was 0.022 (Supplementary Tables S6 and S7).

Table 4.

Model performance comparison in the validation set (n = 4,499)

Model	AUC	N Variables
LACE index (recalibrated)	0.557	4
READMIT-subset (recalibrated)	0.604	6
LASSO + MI + log transforms	0.649	24
Elastic Net + MI	0.649	22
Ridge + MI	0.648	33
Complete-case LASSO	0.656	22

Note. AUC, area under the receiver operating characteristic curve; MI, multiple imputation. DeLong test: LASSO vs. LACE, ΔAUC = 0.092, z = 5.69, p < 0.001; LASSO vs. READMIT-subset, ΔAUC = 0.045, z = 3.47, p < 0.001. Elastic Net, Ridge, and complete-case analyses presented for comparison (Supplementary Table S5)

Fig. 2.

Receiver Operating Characteristic Curves Comparing the Revised LASSO Model with Benchmark Indices. Note. ROC curves comparing the LASSO model with multiple imputation (AUC = 0.649), recalibrated READMIT-subset (AUC = 0.604), and recalibrated LACE index (AUC = 0.557) in the validation set (n = 4,499). DeLong test p < 0.001 for LASSO vs. LACE and LASSO vs. READMIT. AUC, area under the curve

The incremental predictive value of discharge medication variables was assessed by comparing a clinical-only LASSO model (AUC 0.653) against the full model including medications (AUC 0.649). The collective addition of medication variables did not improve discrimination (ΔAUC = − 0.004), indicating that discharge medications do not meaningfully enhance prediction beyond clinical and demographic variables. However, individual medication variables (FGA, LAI, antidepressant) were independently retained by LASSO with small but nonzero coefficients, likely because they serve as markers of illness severity and treatment adherence strategy rather than providing incremental predictive information as a group (Supplementary Table S8).

Clinical risk score

A revised clinical risk score was derived using categorised predictors (Table 5). Integer point values ranged from − 2 to + 5 per predictor, with a total possible score range of − 2 to 17 (Supplementary Fig. S4). Points were assigned as follows: prior admissions in the past 12 months 4+ (+ 5 points), prior lifetime admissions 4+ (+ 3), prior lifetime admissions 2–3 (+ 2), schizoaffective disorder (+ 3), first-generation antipsychotic (+ 2), emergency admission (+ 1), Charlson index ≥ 1 (+ 1), ED visits ≥ 2 in prior 6 months (+ 1), personality disorder (+ 1), neurodevelopmental disorder (+ 1), antidepressant at discharge (+ 1), and long-acting injectable antipsychotic at discharge (− 2 points, protective).

Table 5.

Clinical risk score for 30-day psychiatric readmission

A. Risk Score Points
Predictor			Category	Points
Prior psychiatric admissions (lifetime)			0–1 (reference)	0
			2–3	+ 2
			4+	+ 3
Prior admissions in past 12 months			0–3 (reference)	0
			4+	+ 5
Schizoaffective disorder			Yes	+ 3
First-generation antipsychotic at discharge			Yes	+ 2
LAI at discharge			Yes	−2
Charlson Comorbidity Index			≥ 1	+ 1
ED visits in prior 6 months			≥ 2	+ 1
Emergency admission			Yes	+ 1
Personality disorder			Yes	+ 1
Neurodevelopmental disorder			Yes	+ 1
Antidepressant at discharge			Yes	+ 1
B. Risk Tier Classification
Risk Tier	Score Range	N	30-Day Readmission Rate	95% CI
Low	−2 to 3	8,250	6.7%	6.1–7.2%
Medium	4	2,222	10.0%	8.8–11.3%
High	≥ 5	4,522	18.0%	16.9–19.1%

Note. (A) Integer point values derived from logistic regression coefficients using categorised predictors. The minimum significant coefficient served as the divisor for integer conversion. (B) Observed 30-day readmission rates by risk tier with Wilson score 95% confidence intervals. Risk tiers defined by score tertiles. LAI, long-acting injectable antipsychotic; ED, emergency department; CI, confidence interval

The score model demonstrated discrimination comparable to the full LASSO model (AUC 0.654 vs. 0.649), with minimal bootstrap optimism (0.004), confirming acceptable simplification for clinical use (Supplementary Table S9). Risk tiers based on score tertiles showed clear separation: Low risk (score ≤ 3): 6.7% (95% CI 6.1–7.2%), n = 8,250; Medium risk (score 4): 10.0% (95% CI 8.8–11.3%), n = 2,222; High risk (score ≥ 5): 18.0% (95% CI 16.9–19.1%), n = 4,522. The high-risk tier had 2.7 times the readmission rate of the low-risk tier. Kaplan-Meier analysis confirmed clear separation between risk tiers at both 90-day and 365-day follow-up (log-rank p < 0.001 for both; Supplementary Fig. S5).

Importantly, the risk score does not assign baseline points for marital status. Because marital status was not retained as a meaningful predictor after multiple imputation with married as the reference category, the score functions identically regardless of the completeness of marital status documentation. In a clinical setting with 100% marital status capture, the expected score distribution and tier classification would be unchanged from those reported here: 55.0% of patients in the low-risk tier, 14.8% in the medium-risk tier, and 30.2% in the high-risk tier. This contrasts with a score derived using the unrecorded reference category, which would assign 3–4 baseline points to nearly every patient with recorded marital status, artificially inflating risk classification in settings where marital status is routinely queried.

Secondary and sensitivity analyses

The revised LASSO model demonstrated temporal stability: training on 2018–2023 data and testing on 2024–2025 data yielded an AUC of 0.676 (95% CI 0.651–0.700), indicating robust performance across the study period (Supplementary Tables S12 and S16). The 90-day readmission model showed improved discrimination (AUC 0.685) compared with the 30-day model, while the 7-day model showed reduced discrimination (AUC 0.612). Performance was highest among the depression subgroup (AUC 0.704) and lowest for schizophrenia (AUC 0.618) (Supplementary Table S17). Additional sensitivity analyses excluding short stays (< 48 h) and restricting to the post-COVID period (2020–2025) confirmed model stability (Supplementary Table S12). A first-admission LASSO model, fitted using only variables available at index presentation (excluding all prior utilisation history), yielded an AUC of 0.556 (95% CI 0.513–0.596; Supplementary Table S18). Exploratory addition of interaction terms and alternative machine learning approaches (random forest, gradient boosting) did not meaningfully improve discrimination (maximum ΔAUC + 0.012; Supplementary Table S18).

Competing risks from post-discharge mortality were assessed through quantitative bounding analysis. Based on published 30-day post-discharge psychiatric mortality rates of 0.5–1.5%, the maximum bias in the observed readmission rate was estimated at 0.05–0.16% points, with negligible impact on model discrimination (< 0.003 AUC units; Supplementary Table S13). Individual-level mortality data were not available in the electronic health record extract, precluding formal Fine-Gray competing risks regression.

Decision curve analysis demonstrated net benefit of the local model over treat-all and treat-none strategies across threshold probabilities from 5% to 30%, with consistent superiority over both benchmarks (Supplementary Fig. S2). At a 10% threshold, the model identified 32.6% of patients for intervention with 51.8% sensitivity, 69.6% specificity, and an observed readmission rate of 16.8% among flagged patients (Supplementary Table S19).

Complete-case sensitivity analysis excluding 4,227 encounters with unrecorded marital status (n = 10,767) yielded an AUC of 0.656 (95% CI 0.628–0.685), consistent with the primary analysis and confirming that the imputation approach did not introduce bias.

Discussion

This study represents the first psychiatric readmission prediction model developed and validated in the United Arab Emirates and, to our knowledge, the first such model from the MENA region. Using contemporary penalised regression methods (LASSO) and multiple imputation for missing data, our locally developed model achieved moderate discrimination (AUC 0.649) that significantly outperformed both the LACE index (AUC 0.557) and READMIT-subset (AUC 0.604) in this population. The derived clinical risk score provided clinically meaningful stratification, with the high-risk tier experiencing a readmission rate 2.7 times that of the low-risk tier (18.0% versus 6.7%). Notably, marital status, well established as a predictor in the international literature, was not retained as a meaningful predictor after multiple imputation with married as the reference category, likely reflecting the large proportion of missing data (28.2%) and the systematically different characteristics of the unrecorded group.

Readmission rate in international context

The observed 10.6% thirty-day readmission rate falls below the global pooled estimate of 16% (95% CI: 13–20%) from a meta-analysis [6] and is comparable to rates reported in Hong Kong (7.1%) [3] and some European centres (10–15%) [26, 27], though substantially lower than US Medicare psychiatric facilities (20%) [28]. Contextual factors including population mobility and follow-up pathways may contribute to this finding.

Model performance compared to existing tools

Our model’s AUC of 0.649 aligns with typical psychiatric readmission models (AUC 0.60–0.70) [15], comparable to READMIT’s development c-statistic of 0.631 [13] and Hong Kong’s psychotic spectrum model (0.68–0.71) [3]. The superior performance over both LACE (AUC 0.557), which was never intended for psychiatric populations [11], and READMIT-subset (AUC 0.604), which was derived in Ontario’s publicly funded healthcare system with substantially different patient demographics and service delivery models [13], provides empirical evidence that locally derived models outperform imported tools in this population. The rationale for developing a region-specific model was not the expectation of fundamentally different predictor domains, as the core predictors (prior utilisation, illness severity, comorbidity, emergency presentation) are consistent with the international literature, but rather that the relative importance and coefficient magnitudes of these predictors may differ in a population with distinct demographic characteristics (median age 31, 90% UAE national, 34% substance use disorders), and that no existing tool had been validated in any MENA psychiatric population prior to this study. Specific coefficient differences support this rationale: length of stay, a significant predictor in the READMIT derivation and retained at three of four sites in a recent cross-site transportability study [5], was not selected by LASSO in our cohort, possibly reflecting differences in bed pressure and discharge planning practices. Conversely, discharge medication variables (FGA, LAI), which are absent from READMIT entirely, were independently retained as predictive markers in our population (Table 1). These differences in predictor retention and coefficient structure, rather than the predictor domains themselves, illustrate why locally derived coefficient estimates outperformed imported tools in this setting. The empirical underperformance of both LACE and READMIT in our cohort supports this rationale and underscores the value of local model calibration.

Interpretation of individual predictors

The following interpretations are offered to contextualise the model’s predictor structure within existing clinical knowledge rather than to establish causal relationships. In a prediction model framework, adjusted coefficients reflect the conditional predictive contribution of each variable after accounting for other predictors, but they remain susceptible to unmeasured confounding and should not be interpreted as causal effects [25, 29]. Schizoaffective disorder was associated with higher predicted readmission risk (LASSO coefficient 0.117), consistent with the broader literature identifying schizophrenia-spectrum disorders as having the highest readmission rates among psychiatric diagnoses. Meta-analytic evidence ranks schizophrenia and related psychoses above bipolar disorder, depression, and substance use disorders for 30-day readmission probability [6, 30]. The intermediate nosological position of schizoaffective disorder between schizophrenia and mood disorders may account for this association, as affected individuals combine the relapse patterns of psychotic illness with the episodic course of affective conditions [31].

Marital status was unrecorded for 28.2% of encounters. Missingness was significantly predicted by non-UAE nationality (OR 0.53), fewer prior admissions (OR 0.75), and later admission year (OR 1.07), consistent with a Missing At Random mechanism (Supplementary Table S1). The unrecorded group had a lower readmission rate (7.0%) than patients recorded as single (12.4%) or married (11.1%), likely reflecting shorter system engagement rather than genuinely lower risk. After multiple imputation with married as the reference category, single marital status was not retained as a meaningful predictor (LASSO coefficient 0.005, selected in only 50% of imputed datasets). Although marital status is well established in the psychiatric readmission literature, with meta-analytic evidence supporting an association between single status and readmission (pooled OR 1.52; 95% CI: 1.21–1.91) [6], the absence of an association in the present cohort may reflect the large proportion of missing data and the characteristics of the unrecorded group. Complete-case sensitivity analysis (n = 10,767) yielded consistent results (single LASSO coefficient 0.040). These findings highlight the importance of addressing missing data through appropriate imputation methods rather than retaining unrecorded categories, and suggest that future models should consider incorporating direct measures of social support adequacy.

Long-acting injectable (LAI) antipsychotic prescription at discharge showed a protective association with readmission (LASSO coefficient − 0.094), consistent with observational evidence that LAIs are associated with reduced rehospitalisation in schizophrenia [23] and the adherence-promoting properties of depot formulations. In a prediction model framework, this coefficient may reflect that patients prescribed LAIs represent a subgroup with better treatment engagement and follow-up planning. Nonetheless, the protective signal supports the clinical rationale for considering LAI formulations in patients at elevated readmission risk.

First-generation antipsychotic (FGA) prescription at discharge was the strongest medication-related predictor of readmission (LASSO coefficient 0.138). Most comparative effectiveness studies have found no significant difference in rehospitalisation rates between typical and atypical antipsychotics when controlled for confounding [32]. suggesting this association reflects confounding by indication: patients prescribed FGAs may represent those with treatment-resistant illness, prior poor response to atypical agents, or clinical profiles warranting depot formulations. In a prediction model framework, the FGA variable serves as a marker of illness severity and treatment complexity, consistent with the observation that medication variables collectively did not improve discrimination beyond clinical predictors (ΔAUC = − 0.004).

Prior psychiatric admissions (log-transformed) and emergency department visits (log-transformed) were the strongest predictors in the revised model. The log-transformation captured the nonlinear relationship whereby the incremental risk associated with each additional admission diminishes at higher counts (readmission rate: 0 prior admissions 7.2%, 1 admission 7.4%, 2–3 admissions 12.2%, 4 + admissions 22.5%). This finding is consistent with systematic reviews identifying previous hospitalisation as the most robust predictor of readmission across psychiatric settings [33]. The Charlson Comorbidity Index (LASSO coefficient 0.095) was also retained, consistent with evidence that medical comorbidity is associated with higher psychiatric readmission rates [34].

Discharge medications, when considered collectively, did not provide incremental predictive value beyond clinical and demographic variables (clinical-only model AUC 0.653 vs. clinical plus medications AUC 0.649). This is consistent with prior prediction modelling studies [35] and likely reflects that medication associations with readmission are mediated through adherence, which cannot be measured at discharge [36]. However, LASSO independently retained FGA prescription and LAI prescription as individual predictors with nonzero coefficients, suggesting these variables capture variance in illness severity and treatment strategy not fully explained by other clinical indicators. This distinction, between the collective non-contribution of medication data and the individual predictive signals of specific medication markers, resolves an apparent contradiction and highlights that certain prescribing patterns serve as informative proxies for clinical complexity.

Clinical utility and implementation considerations

The clinical risk score derived from our model demonstrated meaningful stratification, with the high-risk tier (18.0% readmission) clearly differentiated from intermediate (10.0%) and low-risk (6.7%) groups. This approximately 2.7-fold risk differential between highest and lowest tiers is comparable to the READMIT index’s stratification [13] and provides a practical framework for resource allocation. High-risk patients could be prioritised for intensive transitional care interventions, early follow-up appointments, community mental health team referral, or enhanced family psychoeducation, interventions that have demonstrated effectiveness in reducing readmission when targeted appropriately [30].

However, several barriers to implementation warrant acknowledgement. A systematic review found that only one psychiatric prediction model worldwide has undergone formal implementation evaluation studying real-world clinical utility [15]. Successful clinical decision support requires integration into existing workflows without disruption, clinician buy-in through interpretable models, and demonstration that prediction-guided interventions actually improve patient outcomes. Our simple clinical risk score, using readily available variables without requiring complex computation, may facilitate adoption compared to machine learning approaches, though prospective implementation studies remain necessary.

Two diagnostic predictors (personality disorder (coefficient 0.028) and neurodevelopmental disorder (coefficient 0.022)) were retained by LASSO with small coefficients, contributing modest points to the clinical risk score. Their retention reflects the regularised selection process, which shrinks rather than eliminates weak but nonzero predictive signals.

When restricted to first-admission patients only (excluding all prior utilisation history), the model showed substantially reduced discrimination (AUC 0.556, 95% CI 0.513–0.596). Exploratory analyses testing interaction terms (age by diagnosis, diagnosis by medication) and alternative modelling approaches (random forest, gradient boosting) yielded negligible improvements (maximum ΔAUC + 0.012; Supplementary Table S12), confirming that the performance ceiling reflects the informational content of available administrative variables rather than modelling choice. This is consistent with the broader literature: the Chen et al. (2018) Ontario system-level study reported that first-hospitalisation patients had substantially different readmission patterns compared with patients with prior admissions [2], and Ren et al. (2025) demonstrated consistent cross-site performance degradation when prior utilisation features were unavailable [5]. The Meehan et al. (2022) systematic review noted that prior hospitalisation is the most robust predictor across psychiatric prediction models, and models lacking this variable consistently show AUCs in the 0.55–0.60 range [15]. For patients presenting with their index psychiatric hospitalisation, the model offers limited discriminatory value, as their risk profile cannot incorporate the powerful signal provided by previous service utilisation patterns. Alternative approaches incorporating symptom severity measures, treatment response during admission, or social determinants of health merit future investigation [37].

Strengths and limitations

This study has several methodological strengths. It represents the largest psychiatric readmission prediction study from the MENA region (14,994 admissions; 8 years). We followed TRIPOD guidelines, employed split-sample and temporal validation with bootstrap optimism correction, and benchmarked against established instruments (LACE and READMIT) using formal statistical comparisons (DeLong test, NRI, IDI). The clinical risk score was independently validated with minimal discrimination loss (ΔAUC = 0.005). Individual substance types were formally tested rather than simply cited as unavailable, demonstrating negligible incremental value (ΔAUC = + 0.005, 95% CI: −0.003 to + 0.014). Additional methodological strengths include: (1) LASSO penalised regression for predictor selection, aligned with contemporary clinical prediction model guidelines [25]; (2) multiple imputation with 20 datasets for missing marital status data, with formal missingness mechanism assessment supporting the MAR assumption; (3) systematic assessment of nonlinearity for all continuous predictors with log-transformation of count variables; and (4) quantitative bounding analysis for competing risks from post-discharge mortality.

Several limitations warrant consideration. The retrospective single-centre design limits generalisability, and external validation remains necessary before widespread adoption. The single-centre design may overestimate model performance relative to what would be observed in external validation, as prediction models consistently demonstrate performance degradation in external validation. Marital status was unrecorded for 28.2% of encounters. Multiple imputation was used to address this missingness; after imputation with married as the reference category, the apparent marital status effect was substantially attenuated, revealing that the preliminary finding was an artefact of the missing-data reference category. This demonstrates both a limitation of the original dataset and the importance of appropriate missing data methods. Although substance use disorder was captured and individual substance types were available in diagnosis codes, supplementary analysis (Supplementary Table S20; Supplementary Fig. S6) showed that granular substance type modelling added negligible discrimination, though alcohol and opioid use disorders showed paradoxically protective associations warranting further investigation. We lacked data on medication adherence, insight, and social support adequacy. Suicidal ideation was infrequently captured in structured fields (< 0.1%) and could not be evaluated. Our outcome captured readmission to the index hospital only, potentially missing transfers elsewhere. Risk tier cut-points were derived from sample tertiles rather than clinical anchors, a limitation common to prediction model development. The absence of linked mortality data precluded formal Fine-Gray competing risks analysis. However, the 30-day observation window limits mortality-related bias, and quantitative bounding analysis using published mortality rates (0.5–1.5%) demonstrated negligible impact on the observed readmission rate (maximum bias 0.16% points) and model discrimination (Supplementary Table S13). Finally, as a prediction model, the identified associations should be interpreted as predictive markers rather than causal effects. Individual predictor coefficients may be confounded by unmeasured variables, and the adjusted associations should not be used to infer causal mechanisms [29].

Implications for psychiatric care in the UAE and broader Middle East

This study addresses a significant evidence gap for mental health services in the UAE and the broader Middle East [38]. The superior performance of a locally derived model compared with imported Western tools supports the importance of context-specific research in healthcare settings with unique demographic, cultural, and system characteristics. The clinical risk score offers a practical discharge planning tool requiring no additional testing. Identifying the 15–20% of admissions at highest readmission risk enables targeted allocation of limited community mental health resources [20, 39]. The 10.6% 30-day readmission rate provides a baseline benchmark for quality monitoring in UAE psychiatric services [40]. As the UAE continues expanding mental health service infrastructure and workforce capacity, prediction-guided care pathways may support more efficient targeting of follow-up intensity and transitional care [21, 39]. Future research should include external validation in other Gulf countries, prospective implementation studies, and evaluation of interventions targeting high-risk patients. The absence of improved prediction from medication variables underscores the need for better adherence measurement methods in future research.

Conclusion

We developed the first psychiatric readmission prediction model for the United Arab Emirates, demonstrating moderate discrimination that outperformed imported generic and psychiatric-specific indices. The derived clinical risk score provides meaningful stratification that could guide resource allocation for discharge planning and transitional care. While external validation and implementation studies remain necessary, these findings offer an evidence-based foundation for quality improvement in Middle Eastern region psychiatric services and support the value of locally developed prediction tools in healthcare settings with distinctive contextual characteristics.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Author contributions

SAB conceptualized the study, curated the data, developed the methodology, conducted the investigation and formal analysis, and drafted the original manuscript. SFJ provided project administration and resources, supervised the work, and contributed to manuscript review and editing. All authors read and approved the final manuscript.

Funding

None.

Data availability

The electronic health record dataset analysed in this study is not publicly available because it contains sensitive patient information and is subject to institutional governance and legal restrictions. All variables used, outcome definitions, and model parameters required to reproduce the analyses are provided in the manuscript and Supplementary Materials.

Declarations

Ethics approval and consent to participate

This retrospective study used routinely collected electronic health record data. This study was conducted in accordance with the Declaration of Helsinki and ethical approval was obtained from the United Arab Emirates Ministry of Health and Prevention Research Ethics Committee (MOHAP/DXB-REC/No. 91/2024). The requirement for informed consent was waived by the ethics committee because the study involved secondary analysis of existing clinical data and posed minimal risk to participants. All data were de-identified prior to analysis.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.