EmpedocAI: Predicting Migraines and Detecting Early Cognitive Decline

Project Type: Open-source health ML (side project growing into something bigger)

The Problem We're Tackling

Migraines affect over a billion people but get misdiagnosed half the time, and dementia cases are set to triple by 2050. Both conditions cost society massively—migraines alone run up a $78B tab annually in the US. The real kicker? We're still treating these conditions reactively when we could be catching warning signs days in advance.

What We're Actually Building

We're combining wearable sensor data (heart rate, sleep, activity patterns) with local weather conditions and voice analysis to predict migraine attacks up to 5 days out and screen for early cognitive impairment. Our voice models are hitting 0.94+ AUC on cognitive screening, which is pretty solid for something that could run on a phone.

The approach: multimodal transformers that fuse time-series physiological signals with acoustic features. Nothing revolutionary, but applying these techniques to neurology data in an open way hasn't really been done at scale.

Why We Need Compute

Training transformer models on long-sequence physiological data and fine-tuning acoustic models isn't cheap. We're working with time-series that span days and audio processing that needs serious GPU hours.

What Goes Public

Everything. All models go on the Hub with full documentation, training code in notebooks, and we'll release anonymized benchmark datasets where ethics boards allow. We want other researchers to build on this without starting from scratch.

The goal is making clinical-grade prediction tools accessible—not locked behind paywalls or proprietary systems. If someone in an underserved area can run early Alzheimer's screening from their phone, that's worth putting in the hours.

Stack: PyTorch, HF Transformers, Wav2Vec2 for audio, plus custom time-series architectures.