DYNAMIC CLASSIFICATION OF LATENT DISEASE PROGRESSION WITH AUXILIARY SURROGATE LABELS

Area of Science:

• Biostatistics
• Neurology
• Machine Learning

Background:

• Disease progression prediction is challenging due to unknown true disease states and diagnostic limitations.
• Distinguishing Alzheimer's disease (AD) from related dementias like Lewy body dementia (LBD) is difficult without gold-standard diagnoses.
• Existing models often make unrealistic assumptions or suffer from misspecification when using surrogate labels and health markers.

Purpose of the Study:

• To develop a novel statistical model for disease progression prediction using surrogate labels and health markers.
• To improve the accuracy of distinguishing between AD and LBD by addressing limitations of existing generative models.
• To integrate generative and discriminative approaches for robust disease state modeling.

Main Methods:

• Proposed a hybrid model integrating a hidden Markov model (generative) with a time-varying discriminative classification model.
• Developed an adaptive forward-backward algorithm with subjective labels for estimation.
• Utilized modified posterior and Viterbi algorithms for future state prediction using objective markers.

Main Results:

• The proposed adaptive method eliminates the need to model the marginal distribution of longitudinal markers.
• Simulation studies demonstrated significant improvements in finite samples compared to traditional algorithms.
• Analysis of the National Alzheimer's Coordinating Center (NACC) dataset showed improved accuracy in distinguishing LBD from AD.

Conclusions:

• The integrated model effectively handles potentially misspecified surrogate labels and incorporates disease progression markers.
• The novel approach offers a more robust and accurate method for disease progression prediction, particularly in complex neurological conditions.
• This method enhances diagnostic accuracy for conditions like LBD versus AD, aiding clinical decision-making.