ClinFormer: a hybrid deep learning framework combining 1D-CNN, bidirectional LSTM, and Transformer encoder for Alzheimer’s disease classification from structured clinical records

Alzheimer’s disease (AD) is a leading cause of disability in late life, yet gold-standard diagnostic procedures remain costly and inaccessible to most at-risk populations. Reliable digital screening pipelines built on routinely measured clinical variables could extend the reach of early detection. This paper presents ClinFormer, a hybrid deep-learning pipeline that sequentially integrates a one-dimensional convolutional block, a bidirectional long short-term memory encoder, and a Transformer module with multi-head self-attention over a 32-dimensional clinical feature vector. The dataset is a programmatically generated benchmark rather than a primary clinical cohort, and all models were evaluated within it; the reported figures therefore characterise relative architectural behaviour and must not be interpreted as clinical diagnostic accuracy. The model was evaluated on the publicly available Alzheimer’s Disease Dataset (2,149 patients, 34 attributes) using stratified ten-fold cross-validation and synthetic minority oversampling fitted only within training partitions, and compared against three classical machine-learning baselines and three standalone deep-learning ablations. On the held-out test partition ClinFormer achieved AUC-ROC 0.901 (95% CI 0.866–0.935), accuracy 83.9% (79.9–87.9), recall 80.7% (73.1– 87.4) and precision 75.4% (67.5–82.7). Random Forest and XGBoost outperformed ClinFormer on every reported metric 93.8% and 92.9% accuracy against 83.9% and the difference was statistically significant across ten cross-validation folds (paired Wilcoxon p = 0.002, rank-biserial r = −1.00 in both cases. ClinFormer did, however, significantly outperform the support vector machine on both AUC-ROC (p = 0.020) and accuracy (p = 0.014). A regularisation and capacity ablation showed that four of seven interventions improved simultaneously on the baseline train–validation gap of 15.1 percentage points and its test AUC-ROC of 0.899, with the two capacity-reduction variants performing best: halving layer widths gave the highest AUC-ROC (0.913) and removing the second recurrent layer the narrowest gap (11.2 points), indicating that the full architecture is over-parameterised for a dataset of this size. Calibration analysis found ClinFormer the least well calibrated of the four principal models (Brier 0.142 against 0.060 for XGBoost). Shapley explanations computed directly on the model identify a clinically coherent diagnostic failure and demonstrate that local attribution becomes uninformative where the model output saturates.