EarlyDeCog
Fair decisions in neurodegenerative disease early diagnostics
Tackling gender bias in EEG-based AI for Alzheimer's and Parkinson's detection
EarlyDeCog is a R+D project led by Starlab Barcelona, which was awarded in the 2nd Open Call of the European Lighthouse on Secure and Safe AI (ELSA). Running from December 2024 to May 2025, the project addresses a critical blind spot in AI-powered diagnostics: the systematic gender bias that causes AI models to perform significantly worse for women than for men — a disparity with direct consequences for patients whose symptoms are dismissed, delayed, or misclassified.
Neurodegenerative diseases such as Alzheimer's and Parkinson's affect more than 125 million people worldwide and represent a $2.1 billion diagnostic market. Yet current AI models trained on historically male-dominated datasets can carry up to a 10% performance gap against women in Alzheimer's detection and a 1% gap in Parkinson's — subtle but clinically consequential disparities that can mean years of missed or incorrect diagnoses.
Early detection is everything in neurodegeneration. By the time patients receive a confirmed diagnosis, they may already be at an advanced stage of the disease. EarlyDeCog builds on Starlab's proven EEG-based AI platform for neurocognitive screening — a technology capable of detecting signs of cognitive decline up to 8 years before clinical symptoms appear, at 95% accuracy, 77% faster than standard neuropsychological tests, and at 30% lower cost than competing approaches, with no side effects.
The missing piece was fairness. EarlyDeCog set out to integrate rigorous bias mitigation directly into this diagnostic pipeline, ensuring the system works equally well regardless of the patient's sex.
The team compared two bias mitigation strategies. The Equalised Odds post-processing method — which calibrates decision thresholds independently by sex after training — emerged as the preferred approach: it substantially reduced performance gaps between male and female subgroups with minimal loss in overall accuracy. The Exponentiated Gradient Reductions in-processing method delivered stronger parity in some settings but proved less stable on smaller datasets and more computationally demanding. Notably, gender-based feature standardization — an intuitive but naïve fix — was found to be ineffective and sometimes counterproductive.
Key results include an AUC of 0.97 for Parkinson's disease detection (Healthy Control vs. Parkinson's), with fairness-corrected models maintaining accuracy and precision scores above 0.90 across both sexes. In Alzheimer's classification, the fairness-aware pipeline achieved accuracy and precision scores above 0.70 while substantially reducing gender disparity in the model's error rates.