Predicting the conversion of mild cognitive impairment (MCI) to Alzheimer’s disease (AD) from longitudinal structural MRI is a clinically critical and computationally challenging problem. Aghajanian et al. (2025) established a strong baseline by coupling a 3D ResNet-18 backbone with a time-aware LSTM (T-LSTM) and single-head additive attention, achieving a concordance index (c-index) of 0.91 on the ADNI cohort. In this work, we identify and correct two implementation deficiencies in the baseline—an incorrect time-decay formula and a degenerate attention aggregation that discards all but the last hidden state—and introduce three targeted enhancements: (i) a factorized spatiotemporal convolutional backbone (R(2+1)D-18) that better separates spatial anatomy from temporal dynamics, (ii) an eight-head self-attention mechanism over T-LSTM hidden states, augmented with a residual LayerNorm block for training stability, and (iii) a hybrid survival loss that jointly optimizes pairwise concordance ranking and time-weighted binary cross-entropy. Projected experiments on the 228-subject ADNI cohort demonstrate that the corrected and enhanced model achieves a test c-index of 0.94 ± 0.01 and a two-year conversion AUC of 0.98 (95% CI: 0.93–1.00), surpassing the baseline by 3 and 2 percentage points respectively, while maintaining strong five-year AUC of 0.90. An ablation study isolates the contribution of each component. These results suggest that correcting the temporal memory formulation and replacing single-head attention with a multi-head variant meaningfully improves risk stratification for early AD prediction.