Data collection

This single-center retrospective study at Crouse Hospital (Syracuse, NY) collected data from electronic health records (EHR) for patients diagnosed with stroke (ischemic, hemorrhagic) or transient ischemic attack (TIA) who presented to the emergency department (ED) between January 2019 and January 2021. Patient management adhered to the hospital’s clinical protocol for suspected stroke (S1 Fig), which prioritizes timely interventions—including CT imaging, thrombolytic eligibility assessment, and reperfusion therapies. Twenty variables spanning demographic characteristics, clinical presentation, vital signs, laboratory measurements, and neuropsychiatric/substance use indicators were collected for each patient and categorized as outlined in Table 1. These 20 predictors were pre-specified a priori based on clinical relevance, routine availability in the electronic health record during acute stroke evaluation, and consensus among the clinical investigators involved in the study. No separate data-driven feature selection approach was applied before model development, and all pre-specified predictors were retained in the analyses. Discharge disposition was classified into four categories: home, specialized care, home with help, and expired.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Overview of variables used and bivariate analysis. P-values 1, 2, and 3 correspond to comparisons of home vs. specialized care, home vs. home with help, and home vs. expired, respectively. The units for LDL, systolic blood pressure, and blood glucose are mg/dL, mmHg, and mg/dL, respectively.

https://doi.org/10.1371/journal.pone.0353683.t001

Operational definition and timing of predictor capture

Predictor variables were defined using structured EHR data available during the acute hospitalization. Specifically: (1) laboratory values and vital signs reflected the earliest available measurements recorded immediately upon ED triage; (2) admission NIHSS represented the initial neurological assessment scored by the clinical team before treatment initiation; (3) depression and other comorbidities were defined based on prior documented medical history or the active problem list available in the EHR at the time of admission and were not inferred solely from medication proxies or unstructured note abstraction; and (4) last known well (LKW) time was extracted from ED stroke documentation and categorized as <3 hours, 3–6 hours, 6–12 hours, 12–24 hours, and >24 hours. Cases with undocumented LKW times (Table 1: n = 7, 0.4%) were treated as missing and handled using multiple imputation by chained equations (MICE), as detailed below.

Missing data handling

Missing data were present in seven variables, ranging from 0.4% to 9.6%, as shown in Table 1. Missing values were handled using a mixed-type MICE-style iterative imputation procedure implemented with IterativeImputer. The imputation procedure was incorporated within the modeling pipeline to reduce information leakage. Thirty imputation iterations were used. This value was selected as a conservative upper limit for the chained-equation iterations, given the relatively low proportion of missingness in the dataset, and to allow sufficient stabilization of the imputation process before downstream model training and evaluation.

Exploratory principal component analysis

To provide an exploratory visualization of the overall structure of the dataset, we performed a three-dimensional principal component analysis (PCA) using all 20 clinical variables. PCA was used solely for descriptive visualization of the degree of overlap and relative organization of discharge categories in reduced-dimensional space and was not used for feature selection, model training, or model evaluation.

Bivariate analysis

We conducted bivariate analyses to assess baseline associations between discharge outcomes, using “home” as the reference category. Chi-square tests evaluated relationships between categorical features and discharge categories, substituting Fisher’s exact test when cell frequencies were below five. Numerical variables were initially tested for normality using Shapiro-Wilk tests, and comparisons across discharge groups were performed with one-way analysis of variance (ANOVA), for normally distributed data or Kruskal-Wallis tests, for non-normal data.

Artificial intelligence analysis

We employed AI-based predictive modeling, defining the target variable (discharge disposition) as a multiclass outcome with four distinct categories: home (class 0), specialized care (class 1), home with help (class 2), and expired (class 3). The predictive features included 20 variables related to demographic information, clinical presentation, treatment, vital signs, laboratory measurements, and neuropsychiatric/substance use factors, as outlined in Table 1. The dataset was randomly partitioned into training (80%) and testing (20%) subsets. Ten classifiers were utilized to predict discharge status: logistic regression, support vector classifier (SVC), decision tree, k-nearest neighbors (KNN), random forest, adaptive boosting, extreme gradient boosting (XGBoost), categorical boosting, a stacked ensemble (random forest + XGBoost), and multilayer perceptron (MLP). A stacked ensemble combining random forest and XGBoost was included as an exploratory hybrid benchmark because these two tree-based methods offer complementary strengths for structured clinical data. Random forest is variance-reducing and robust through bagging and random feature selection, whereas XGBoost can reduce bias and capture complex nonlinear relationships through boosting. The ensemble was implemented using a logistic regression meta-learner trained on the predicted probabilities of the two base models after individual hyperparameter tuning.

Multiclass AI model development, tuning, and evaluation

All classifiers were trained using their standard multiclass implementations for the four-class discharge outcome. Most classifiers directly support multiclass classification, whereas scikit-learn’s SVC uses a native one-vs-one (OvO) strategy for multiclass training. We did not impose a common one-vs-rest training wrapper across all models. For class-specific performance evaluation, including receiver operating characteristic (ROC) curve / area under the ROC curve (AUC) and calibration analyses, a one-vs-rest (OvR) framework was used for each discharge category. This approach provides clinically interpretable, destination-specific estimates by comparing each outcome against all remaining outcomes. For the SVC, the OvR evaluation was performed using the final class-probability outputs and did not alter the underlying one-vs-one training strategy.

The final selected MLP classifier was implemented using scikit-learn. The tuned network consisted of a single hidden layer with 48 neurons and a logistic activation function, with an output layer generating class probabilities for the four discharge disposition categories. The model was trained using the Adam optimizer with a constant learning rate of 0.0007, mini-batch size of 8, and a maximum of 3000 training iterations. No dropout or batch normalization layers were used. Regularization was instead performed using L2 weight decay, with a final selected alpha value of 0.03, together with early stopping. Early stopping was enabled with a tolerance of 1 × 10^-4 and n_iter_no_change = 35. For early stopping, an internal validation fraction of 0.25 was used. During cross-validation, this internal validation subset was drawn only from the training portion of each fold. No independent hold-out test data were used for early stopping, hyperparameter tuning, model selection, probability calibration, or threshold determination.

Given the class imbalance—home (44.9%), specialized care (29.5%), home with help (18.3%), and expired (7.3%)—we incorporated class-imbalance handling into model training. Native class weighting was used when supported, whereas weighted-resampling fallback approximated class-weighted learning for the MLP. No synthetic oversampling method, including synthetic minority over-sampling technique (SMOTE), was applied [28]. Hyperparameter tuning was performed exclusively within the training set using stratified 5-fold cross-validation. Multiple performance metrics were evaluated during cross-validation; however, the primary selection and refit metric was the mean cross-validated macro-F1 score, which was selected a priori because the outcome was multiclass and class-imbalanced. The best-performing classifier was selected based on the highest mean cross-validated macro-F1 score across the validation folds. Accuracy, AUROC, sensitivity, and specificity were additionally examined as secondary performance measures but did not alter the predefined model-selection criterion. The full hyperparameter search spaces and final selected settings for all classifiers are provided in S1 Table. After hyperparameter selection, probability calibration of the final selected MLP classifier was performed using vector scaling fitted on out-of-fold predictions from the training set only. Class-specific threshold tuning was also performed using calibrated out-of-fold predictions from the training set only. The independent hold-out test set was strictly isolated from hyperparameter tuning, model selection, probability calibration, and threshold determination. The hold-out test set was used only after final model selection for final performance evaluation and post hoc SHAP interpretation. Calibration performance was assessed on the independent hold-out test set using class-wise one-vs-rest reliability diagrams. For each discharge category, calibration performance was summarized using the Brier score and expected calibration error (ECE). Ten bins were used for each class-wise calibration curve. Learning curve analysis was performed to evaluate model stability by plotting training and cross-validation macro-F1 scores across increasing training-set sizes. Generalization was further validated by comparing cross-validation metrics to hold-out test performance, ensuring minimal divergence.

SHAP analysis

To transparently identify key predictors of discharge disposition and ensure interpretability, we conducted a systematic SHAP analysis on the best-performing classifier. The process comprised three rigorous steps [27]:

1. - Step 1: Feature Ranking and Subset Aggregation:
   1. Class-Specific SHAP Importance: For each discharge class (home, specialized care, home with help, expired), features were ranked by descending SHAP importance values. SHAP importance reflects the average absolute impact of a feature on model predictions across all instances.
   2. Threshold-Based Subsets: From each ranked list, the top 65%, 70%, 75%, 80%, and 85 of features were iteratively selected. Subsets from all four classes were then aggregated into a single union set, ensuring comprehensive coverage of predictors relevant to all discharge outcomes.
2. - Step 2: Iterative Model Retraining and Threshold Selection
   1. Retraining on Subsets: The best-performing model (with original hyperparameters) was retrained on each aggregated subset (65%, 70%,  ..., 85%).
   2. Performance Evaluation: The performance of the classifier, trained on each specific subset of features using the original hyperparameters, was compared to that of the full-feature model using accuracy, macro-averaged F1-score, and AUC to identify the optimal feature subset.
3. - Step 3: Sensitivity Analysis for Robustness Validation:
   1. To validate stability, features were individually added to or removed from the optimal subset. A subset was considered robust if perturbation (adding or removing any single feature) resulted in minimal alteration in accuracy, F1-score, and AUC values, with values remaining close to those of the original full-featured model.

The supplementary analysis code and serialized model-related artifacts are provided as S1 Code.

Statistical analysis

Bivariate analysis revealed numerous significant associations between baseline variables and discharge disposition when compared to the reference category (discharge to home). Specifically, 12 predictors were associated with discharge to specialized care, 10 with discharge to home with help, and 14 with in-hospital mortality (expired) (Table 1). While informative for initial assessment, bivariate analyses do not account for the complex interdependencies between variables.

Explainable AI analysis

According to the stratified 5-fold cross-validation results on the training set, the MLP classifier achieved the highest mean cross-validated macro-F1 score among the ten evaluated classifiers and was therefore selected as the best-performing model for subsequent calibration and SHAP analyses. On the independent hold-out test set, the selected MLP achieved an accuracy of 0.646, with macro-averaged specificity, precision, sensitivity, and F1-score values of 0.873, 0.559, 0.557, and 0.548, respectively (Table 2). The class-wise AUC values were 0.901 for Home, 0.874 for Specialized Care, 0.674 for Home with Help, and 0.889 for Expired (Fig 2a, 2b). Robustness assessments indicated excellent model stability and generalization: discrepancies between training and testing accuracy were minimal, differences between test and cross-validation accuracy were less than 4.6%, learning curve analysis showed a negligible rightmost gap (4.7%; Fig 2c), and cross-validation accuracy exhibited low variability (standard deviation = 0.023; Table 2). These findings collectively suggest that the model was stable and generalized well, with a low apparent risk of overfitting or underfitting. In addition, calibration analysis of the final selected MLP classifier demonstrated reasonable, although class-varying, agreement between predicted probabilities and observed event frequencies across the four discharge categories (Fig 2d). The corresponding Brier score / ECE values were 0.130 / 0.060 for home, 0.133 / 0.069 for specialized care, 0.142 / 0.020 for home with help, and 0.048 / 0.018 for expired.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Comparison of the performance of ten classifiers using overall test-set metrics, cross-validation accuracy, and class-wise AUC values. Cross-validation accuracy ± SD represents the mean cross-validation accuracy with standard deviation across folds. The area under the receiver operating characteristic curve (AUC) 1, 2, 3, and 4 correspond to Home, Specialized Care, Home with Help, and Expired classes, respectively.

https://doi.org/10.1371/journal.pone.0353683.t002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Performance evaluation of the final MLP classifier.

(a) Class-wise one-vs-rest ROC curves with AUC values for home, specialized care, home with help, and expired discharge outcomes. (b) Percentage-based confusion matrix showing predicted versus true discharge categories. (c) Learning curve showing training and cross-validation macro-F1 across increasing training-set sizes. (d) Class-wise one-vs-rest calibration curves for the MLP classifier, with Brier scores and ECE values indicating agreement between predicted probabilities and observed outcome frequencies.

https://doi.org/10.1371/journal.pone.0353683.g002

SHAP analysis of the optimized MLP (the best-performing classifier) provided insights into feature contributions for predicting each discharge category. The results indicated that the optimal subset of features comprised the top 80% (S2 Table). Specifically, for predicting discharge to home (class 0), the key predictors by SHAP importance, ranked in descending order, were length of stay (LOS), admission NIHSS, age, primary diagnosis, sex, blood glucose, aspirin use, and depression (Fig 3a). For specialized care (class 1), the key features, ranked by SHAP importance, included LOS, primary diagnosis, admission NIHSS, age, anticoagulant, insurance type, LKW, and depression (Fig 3b). The prediction of home with help (class 2) was most influenced by primary diagnosis, admission NIHSS, LOS, sex, age, aspirin use, insurance type, and blood glucose (Fig 3c). Finally, for expired (class 3), the most critical predictors, ranked by SHAP importance, were admission NIHSS, primary diagnosis, age, LOS, anticoagulant, troponin, smoking status, and tPA administered (Fig 3d). It should be noted that the sensitivity analyses presented in S2 Table confirmed the robustness of the identified key features (Table 2 and S2 Table).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. SHapley additive exPlanations (SHAP) summary plots for discharge disposition predictions.

Class-specific SHAP values quantify feature contributions to the MLP model’s predictions: (a) Home discharge (Class 0), (b) Specialized care (Class 1), (c) Home with help (Class 2), and (d) Expired (Class 3).

https://doi.org/10.1371/journal.pone.0353683.g003

Potential clinical implementation of the explainable AI framework

The class-specific SHAP findings highlight the multifactorial nature of discharge disposition after acute stroke and support the clinical plausibility of the model’s predictions. Across all four discharge categories, LOS, admission NIHSS, age, and primary diagnosis emerged as shared key predictors, indicating that the model consistently integrated hospitalization trajectory, neurological injury severity, physiologic reserve, and stroke subtype when estimating discharge destination. Importantly, the SHAP-derived age threshold of 72 years provides a clinically useful marker for early admission-based discharge-risk stratification. Although age itself is not modifiable, identifying patients above this threshold may help clinicians anticipate reduced likelihood of independent home discharge and increased need for specialized care, home-based assistance, or mortality-risk monitoring early in the hospitalization.

Beyond these shared predictors, several SHAP-highlighted variables identify modifiable or addressable care processes. The contribution of blood glucose and LKW time emphasizes the importance of early metabolic optimization, rapid stroke recognition, and timely activation of acute treatment pathways. Similarly, depression highlights the need for early mental health screening and psychosocial assessment, as neuropsychiatric burden may reduce rehabilitation engagement, increase caregiver strain, and delay safe discharge planning. Insurance type, which contributed to the prediction of home with help and specialized care, further underscores that discharge disposition is shaped not only by neurological severity, but also by access to post-acute resources, payer authorization, home health availability, and rehabilitation placement logistics.

In an EHR-integrated workflow, patients predicted to be discharged home with help could trigger early social work involvement, case management, caregiver assessment, home health coordination, and optimization of comorbidities such as dysglycemia or depression. Patients predicted to require specialized care could prompt earlier rehabilitation consultation, therapy planning, insurance authorization, and communication with post-acute facilities. Patients at elevated risk for expired status could prompt intensified neurological stabilization, cardiopulmonary monitoring, and goals-of-care discussions when clinically appropriate. Thus, the proposed explainable AI framework may support earlier and more individualized care planning by linking predicted discharge trajectories to transparent SHAP-based explanations, while complementing rather than replacing clinician judgment.

Limitations and future directions

This single-center retrospective study has several limitations. Although the dataset included routinely available clinical variables, it lacked more granular predictors, such as detailed neuroimaging characteristics, precise infarct or hemorrhage volume, lesion location, vascular territory involvement, edema burden, and markers of mass effect or secondary hydrocephalus. Prior computational studies relating cerebrospinal fluid hydrodynamic alterations to clinical symptoms and treatment response in neurosurgical diseases have demonstrated the prognostic value of patient-specific biomechanical data [38–40]. Extending this paradigm to stroke discharge disposition, future AI frameworks could integrate advanced numerical and fluid–structure interaction markers to capture the mechanical impact of stroke-related edema, hemorrhage, or secondary hydrocephalus. In addition, the complex and context-dependent role of anticoagulant use warrants targeted investigation in specific clinical settings such as atrial fibrillation management, and the prognostic implications of specific hemorrhagic stroke subtypes (e.g., subarachnoid, intracerebral, and subdural hemorrhage) were not analyzed separately, limiting our understanding of their distinct effects on discharge disposition. Moreover, the focus on acute hospitalization data did not capture factors such as ongoing recovery potential and post-discharge rehabilitation quality, which are important determinants of longer-term outcomes and should be incorporated into future longitudinal models. Heterogeneity across acute stroke categories, including etiologic ischemic stroke subtypes such as cardioembolic, lacunar, atherothrombotic, unusual-etiology, and undetermined-etiology infarction, may also have influenced outcome prediction but was not explicitly modeled in the present study. Finally, although the current AI framework demonstrated favorable internal performance, it should still be regarded as a feasibility proof-of-concept. External validation across multiple institutions is required before development of a clinical prototype or implementation in bedside decision-making. Potential effects on discharge efficiency, resource allocation, and health-system outcomes should be evaluated prospectively before clinical deployment.