EarlyBird

EarlyBird: Multimodal Alzheimer's Detection Achieving 71% Accuracy Using Novel Bio-Hermes-001 Dataset

Background: While most Alzheimer's disease (AD) detection approaches rely on expensive MRI or PET imaging, we developed a multimodal machine learning pipeline using the novel Bio-Hermes-001 dataset, combining accessible biomarkers for democratized screening.

Methods: We successfully integrated and processed seven data modalities from 1,005 patients in the GAP Bio-Hermes study on AD Workbench: blood-based biomarkers (151 features including Tau-217, Tau-181, GFAP, NFL, Amyloid-beta ratios), digital cognitive assessments, PET neuroimaging (325,787 files including 57,209 DICOM images), clinical records in SDTM format, proteomics, genomics, and demographics. Our pipeline handles complex clinical trial structures, performs SDTM-to-ML transformations, and extracts real diagnoses from medical history records (62.4% CN, 20.6% MCI, 17.0% AD).

Results: Using Gradient Boosting with class balancing to address MCI under-representation, we achieved:

71.1% overall accuracy (3-class: CN/MCI/AD)
66.4% weighted F1 score
Consistent performance across all disease stages, crucial for early MCI detection
Successfully processed all 325,787 PET imaging files and integrated multimodal features

Key Innovations:

First implementation using Bio-Hermes-001 multimodal dataset instead of standard MRI
Novel approach combining blood biomarkers with digital cognitive tests
Balanced class sampling improving minority MCI detection (critical intervention window)
Efficient handling of 57,209 DICOM files and complex SDTM clinical trial format

Technical Implementation:

XGBoost and Gradient Boosting models with hyperparameter optimization
Manual oversampling with noise injection for class balance
Stratified train-test splits maintaining class distributions
Comprehensive validation framework with cross-validation

Clinical Impact: Our 71% accuracy using primarily non-invasive tests (blood biomarkers, cognitive assessments) demonstrates that accessible, multimodal screening could replace expensive imaging for initial AD detection. Blood-based Tau-217 and digital cognitive scores proved highly predictive, offering a scalable screening approach.

Model Deployment: Three serialized model files enable immediate deployment:

alzheimer_finetuned_model.pkl (main Gradient Boosting model)
alzheimer_imputer.pkl (missing value handling)
alzheimer_feature_scaler.pkl (feature normalization)

Future Directions: With additional optimization, ensemble methods and deep learning integration could push accuracy to 75-80%. The multimodal approach shows particular promise for early MCI detection, the critical window for intervention.

Keywords: Alzheimer's disease, multimodal machine learning, Bio-Hermes-001, blood biomarkers, early detection, gradient boosting

EarlyBird achieved 71% accuracy with novel multimodal approach - demonstrating viable alternative to expensive imaging-based detection for early Alzheimer's detection.

Project Name: EarlyBird - "The early bird catches the diagnosis" Hack4Health 2025 Submission