Hybrid intelligent decision support system based on CatBoost, SHAP analysis, and fuzzy logic for assessing Alzheimer's disease risk using clinical data

The growing prevalence of Alzheimer's disease and the limited capacity of specialized neuropsychological diagnostics create a need for automated tools for preliminary risk assessment suitable for use at the primary outpatient screening stage. This study proposes a three-stage hybrid model combining CatBoost gradient boosting, the SHAP explainable artificial intelligence method, and a Mamdani-type fuzzy inference system. The classifier was trained on a structured dataset of clinical, demographic, behavioral, and cognitive indicators from 2149 patients (32 features after excluding identifiers) and demonstrates an accuracy (accuracy) of 0.9535, a recall of 0.9276, an F1-score of 0.9338, and an area under the ROC curve of 0.9444 on an independent test set. SHAP analysis revealed a pronounced dominance of indicators of functional and cognitive capacity (Functional Assessment, ADL, Memory Complaints, MMSE, Behavioral Problems), with a negligible contribution from some clinical-laboratory and behavioral features, which is broadly consistent with the clinical logic of establishing a diagnosis. A fuzzy decision support component was constructed based on the classifier's output probability and the two most informative functional features, converting a point probability estimate into a three-level risk scale (low/medium/high) with an area under the ROC curve of 0.943, which ensures interpretability of the result for the physician while preserving the discriminative ability of the original model. It is shown that combining accurate but poorly transparent boosting with a transparent fuzzy layer creates a workable compromise between accuracy and interpretability in clinical decision support systems. The growing prevalence of Alzheimer’s disease and the limited throughput of specialised neuropsychological diagnostics call for automated risk-screening tools suitable for primary outpatient use. This paper proposes a three-stage hybrid model combining Cat Boost gradient boosting, the SHAP explainableAI method, and a Mamdani-type fuzzy inference system. The classifier was trained on a structured set of clinical, demographic, behavioural, and cognitive indicators for 2,149 patients (32 predictors after identifier removal) and, on an independent test set, achieves an accuracy of 0.9535, a recall of 0.9276, an F1-score of 0.9338, and a ROC-AUC of 0.9444. SHAP analysis revealed a pronounced dominance of functional and cognitive indicators (Functional Assessment, ADL, Memory Complaints, MMSE, Behavioral Problems), while several clinical-laboratory and lifestyle features contributed negligibly, which is broadly consistent with the clinical logic underlying the diagnosis. Using the classifier’s output probability together with the two most informative functional features, a fuzzy decision-support block was built that converts a point probability estimate into a three-level risk scale (low/medium/high) with a ROC-AUC of 0.943, providing an interpretable output for the clinician while preserving the discriminative power of the underlying model. The results show that combining an accurate but opaque boosting model with a transparent fuzzy layer offers a workable trade-off between accuracy and interpretability for clinical decision-support systems.