Modeling multiple-criterion diagnoses by heterogeneous-instance logistic regression.

Area of Science:

• Statistics
• Medical Informatics
• Neuroscience

Background:

• Mild cognitive impairment (MCI) is a precursor to Alzheimer's disease (AD), posing significant caregiving and economic burdens.
• Current MCI/AD diagnosis relies on cognitive domain impairment but often lacks specific domain status in medical records, treating them as missing data.
• Traditional multiple-instance learning methods are unsuitable due to differing predictors across cognitive domains.

Purpose of the Study:

• To develop a novel statistical model for diagnosing MCI and AD that accommodates heterogeneous predictors across cognitive domains.
• To address the challenge of missing cognitive domain status information in medical records.
• To provide accurate estimation and prediction for MCI/AD diagnoses.

Main Methods:

• Generalized multiple-instance logistic regression to create a heterogeneous-instance logistic regression model.
• Employed the expectation-maximization algorithm for parameter estimation due to missing variables.
• Developed specific model variants for MCI and AD diagnoses.

Main Results:

• Validated the proposed model's estimation accuracy, latent status prediction, and robustness through extensive simulations.
• Demonstrated the model's practical utility by analyzing the National Alzheimer's Coordinating Center-Uniform Data Set.
• The model effectively handles missing domain status data and varying predictors.

Conclusions:

• The proposed heterogeneous-instance logistic regression model offers a robust and accurate approach for MCI/AD diagnosis.
• This method improves upon traditional approaches by handling domain-specific predictors and missing data.
• The model shows significant potential for clinical application in diagnosing neurodegenerative diseases.