In-depth insights into Alzheimer's disease by using explainable machine learning approach

Area of Science:

• Medical Informatics
• Computational Neuroscience
• Machine Learning in Healthcare

Background:

• Early diagnosis of Alzheimer's disease (AD) remains challenging due to its complexity and the late onset of cognitive symptoms.
• Existing diagnostic methods often fall short, necessitating advanced analytical approaches for timely detection.
• Large-scale datasets combining medical, cognitive, and lifestyle information offer potential for uncovering early AD indicators.

Purpose of the Study:

• To analyze a comprehensive dataset of over 12,000 individuals to identify early indicators of Alzheimer's disease.
• To develop and interpret an explainable machine learning model for improved diagnostic accuracy.
• To validate or refute established hypotheses regarding AD risk factors through data-driven insights.

Main Methods:

• Data preprocessing techniques including handling missing data, redundancy reduction, and correlation analysis were applied.
• An XGBoost model was trained and optimized, with hyperparameter tuning and evaluation using a F1-score of 0.84.
• Explainability was achieved using SHAP (SHapley Additive exPlanations) values for global and local model interpretation.

Main Results:

• The XGBoost model achieved a competitive F1-score of 0.84, demonstrating high performance in AD prediction.
• SHAP analysis provided a clear scheme illustrating the positive and negative influences of various features on AD diagnosis.
• Data-driven interpretability challenged and confronted previously established hypotheses about AD.

Conclusions:

• Explainable machine learning approaches are crucial for demystifying complex diagnostic processes like early Alzheimer's detection.
• The derived interpretability scheme offers valuable insights for physicians and experts in diagnosing early-stage AD.
• This research highlights the power of integrating large datasets with interpretable AI for advancing neurological disease diagnostics.