AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Decoding the Hypotensive Tipping Point? A Time-Series Transformer for Dynamic Prediction of Incident Cognitive Impairment in SPRINT-MIND — And a Benchmark Against Classical Models
By Bolaji O, Nyirenda N, Bandaru RR et al.
Background: Whether visit-to-visit trajectories of blood pressure and labs—versus single measurements—improve prediction of cognitive impairment after intensive BP lowering is unknown. We built a time-series survival Transformer generating a continuously updating ("living") risk score for amnestic MCI or probable dementia in SPRINT/SPRINT-MIND, benchmarked against classical and ML models. Methods: From 8,563 participants (871 composite events), we built a leakage-free longitudinal tensor. Each encounter carried SBP/DBP, heart rate, pulse and mean arterial pressure, eGFR, sodium, potassium, antihypertensive class count, and inter-visit interval. A discrete-time survival Transformer was compared with a GRU, time-varying Cox, landmark Cox, an age-plus-baseline-cognition Cox, and Random Survival Forest/gradient-boosted models, all scored under one landmark protocol (Uno's C-index, dynamic AUC, Brier). Results: Event counts matched the trial (dementia 325, MCI 640, composite 871). Current eGFR was independently associated with the outcome (HR 0.97 per 10 mL/min, p=0.001); current and baseline SBP were not. All models showed near-identical discrimination (Uno C≈0.86-0.87, overlapping CIs; Transformer 0.865 vs age+cognition Cox 0.865). Integrated gradients showed risk dominated by baseline cognition and age—~100х any longitudinal channel. Highest-risk patients had lower eGFR and higher, more variable SBP, not a hypotensive signature. Conclusions: A state-of-the-art sequence model matched but did not exceed a two-variable clinical model. The longitudinal hemodynamic stream adds little beyond age and baseline cognition, refuting a simple "hypotensive tipping point." The model's value is continuous with transparent re-estimation.