Early identification of advanced chronicity (MACA) patients using Machine Learning models: a population-based predictive approach for proactive care stratification

Technical Summary: Early Identification of MACA Patients Using Machine Learning

Problem Statement
The identification of patients with Advanced Chronic Conditions (MACA) remains a critical challenge in clinical practice. Current strategies often rely on retrospective criteria or clinical judgment (e.g., NECPAL tool, Charlson Comorbidity Index), which can delay timely, personalized intervention. These traditional methods face limitations in sensitivity for early-stage identification and struggle to capture the multidimensional, dynamic nature of advanced chronicity, particularly in non-oncological populations where prognostic uncertainty is high. While Electronic Health Records (EHRs) offer vast data potential, there is a notable gap in studies specifically targeting the early, proactive detection of MACA patients in real-world, population-based settings using Machine Learning (ML).

Methodology
This study employed a retrospective observational cohort design at Hospital Universitario Parc Taulí (Sabadell, Spain) to develop and internally validate an ML-based predictive model.

• Data Source & Population: The dataset comprised 163 patients (80 MACA, 83 non-MACA) extracted from hospital EHRs. Exclusion criteria applied to Pediatrics, Obstetrics, Gynaecology, Mental Health, and Intensive Care units to ensure clinical coherence with the definition of advanced chronicity.
• Feature Engineering: Initially, 173 variables across 24 clinical conditions were identified. Due to extraction constraints, 80 variables were selected. Crucially, variables reflecting prior engagement with specialized palliative care (Condition 1) were excluded to prevent circular dependencies and ensure the model identifies previously unrecognized cases. Hierarchical dependencies between related conditions (e.g., frailty progression) were explicitly incorporated.
• Preprocessing: Missing values were handled via mean imputation (numerical) or zero-assignment (categorical). Categorical variables underwent one-hot encoding, and continuous variables were standardized.
• Feature Selection: Three strategies (k-best, low variance, recursive feature elimination) reduced the dataset to 10 key predictors.
• Model Development: 14 supervised classification algorithms were evaluated, including linear models (Logistic Regression, LDA, SVM), probabilistic methods (Naïve Bayes), and ensemble methods (Bagging, Random Forest, XGBoost, LightGBM, CatBoost, AdaBoost).
• Validation Strategy: A dual validation approach was used: a 75/25 train-test split and a stratified 5-fold Cross-Validation (CV). Given the goal of early detection, sensitivity (recall) was prioritized over specificity to minimize false negatives.

Key Results

• Feature Selection: The final model utilized 10 predictors: absolute dependency (COND8), severe dependence/functional decline (COND9), moderate dependence (COND10), use of specific resources (COND19), hard/medium geriatric syndromes (COND22/23), complex chronic patient identifier, outpatient palliative assessment, home care intervention, and neuroleptic prescription.
• Model Performance:
  • Bagging Classifier: Achieved the most consistent performance with a sensitivity of 0.91, precision of 0.905, F1-score of 0.905, and an AUC-ROC of 0.902 on the test set.
  • Decision Tree: Performed identically to Bagging on the test set (Accuracy, Precision, Recall, F1-score all at 90.5%; AUC 0.912).
  • Cross-Validation: The 5-fold CV confirmed model stability. Bagging showed superior robustness with a mean F1-score of 0.919 and mean AUC of 0.953, compared to 0.891 and 0.893 for the Decision Tree, respectively.
• Predictor Importance: Feature importance analysis on the Bagging model identified "absolute dependency" (weight 0.318) and "use of specific resources" (weight 0.262) as the most influential predictors.

Key Contributions

1. Proactive Detection Framework: The study demonstrates that ML models can effectively identify MACA patients using routine EHR data, shifting from reactive, event-triggered identification to proactive screening.
2. Parsimonious & Interpretable Model: The research establishes that a reduced set of 10 clinically interpretable variables (focusing on functional decline, dependency, and resource utilization) is sufficient for high-discriminative performance, facilitating clinical integration.
3. Avoidance of Circular Logic: By excluding variables related to prior palliative care identification, the model is designed to detect unrecognized cases rather than simply reproducing existing clinical decisions.
4. Robust Ensemble Performance: The study validates that ensemble methods, specifically Bagging, offer a favorable balance between sensitivity and robustness for this specific clinical task, outperforming simpler linear models and showing greater stability in cross-validation than single Decision Trees.

Significance and Claims
The paper claims that its findings highlight the potential of ML-based tools for the early detection of advanced chronicity, offering a high discriminative capacity (AUC > 0.90) that could support proactive care stratification. The authors assert that the model's ability to utilize routinely available data, combined with the interpretability of its predictors, makes it a viable decision-support tool for shifting healthcare delivery toward a more efficient, patient-centered framework.

However, the authors maintain a modest tone regarding the study's limitations. They explicitly acknowledge that the small sample size (n=163) constrains generalizability and increases the risk of overfitting, despite internal validation. They note that the near-balanced class distribution does not reflect real-world prevalence, which may lead to miscalibration in practice. Consequently, the paper concludes that while the results are promising, external validation in independent, multicenter cohorts and prospective clinical studies are essential to confirm the model's stability, transportability, and actual impact on patient outcomes and resource allocation.