A Machine Learning-Based Framework for differential screening of Neurodegenerative Dementias using routine clinical data

Background The increasing prevalence of neurocognitive disorders (NCDs), including Alzheimer’s disease (AD), vascular dementia (VaD), dementia with Lewy bodies (DLB), primary progressive aphasia (PPA), and behavioral variant frontotemporal dementia (bvFTD), highlights the need for accessible tools to support early diagnosis. Machine learning (ML) may enhance diagnostic accuracy by leveraging routinely collected clinical information. We evaluated an ML pipeline based exclusively on first-visit neurological assessments to support differential diagnosis of major NCDs. Methods We retrospectively analysed clinical data from 932 patients. A CatBoost multiclass classifier was trained using 43 routinely available variables, including demographic characteristics, neurological signs, and symptom chronology. Bayesian optimization (Optuna) was used to maximize the macro-averaged F1-score while addressing class imbalance. Model interpretability was assessed through SHAP (SHapley Additive exPlanations), enabling patient-level and global feature attribution. Performance was evaluated on a held-out test set using accuracy, precision, recall, and F1-score. Results The model achieved 73.8% overall accuracy, with a macro F1-score of 0.646 and a weighted F1-score of 0.743. Diagnostic performance was highest for AD (F1 = 0.821), followed by DLB (0.704), bvFTD (0.681), PPA (0.667), and VaD (0.357). SHAP analysis identified first symptom, age at diagnosis, and language symptom chronology as the most influential predictors, closely reflecting established clinical diagnostic criteria. Conclusions Symptom chronology emerged as a robust diagnostic discriminator, underscoring the clinical value of temporal disease trajectories. The prototype Streamlit interface demonstrated practical clinical deployment by providing probabilistic predictions with transparent explanations. These findings support clinically driven, biomarker-free ML models as scalable first-line decision-support tools, particularly in resource-limited settings, while emphasizing the need for prospective multicenter external validation before clinical implementation.