Machine Learning Approaches for Cognitive Disorder Detection using Administrative Health Data

Background: Cognitive disorders, like mild cognitive impairment (MCI) and dementia, are major public health issues, yet early detection is often delayed. Administrative health data offers a potential resource for early detection and population-level surveillance of cognitive disorders, but traditional rule-based algorithms and standard statistical approaches have poor performance with MCI detection and differentiating severity of cognitive disorders. Machine learning (ML) may offer improved capabilities compared to previous approaches. Objective: This thesis aimed to develop and evaluate ML algorithms using linked administrative health data (AHD) and specialist clinical data to 1) distinguish individuals with no cognitive disorders (NCD) from dementia and NCD from MCI, and 2) perform multiclass classification (NCD vs. MCI vs. dementia) against specialist diagnoses. Methods: This cross-sectional study linked PROMPT specialist diagnoses (reference standard) with up to 5 years of AHD from Alberta physician claims, hospitalizations, emergency visits, and prescriptions for 1,789 individuals between 2010-2023. Six ML models (logistic regression, random forest, gradient boosting, support vector machine, extreme gradient boosting, stacking ensemble) were trained and evaluated, primarily using 'definite' diagnoses as the cognitive outcomes. Performance was assessed using balanced accuracy, using bootstrapping to compute confidence intervals. Permutation feature importance was used to identify key predictors. Results: ML models achieved reasonable performance differentiating NCD from dementia with a balanced accuracy up to 0.86 (95% CI: 0.79-0.92). Performance was reduced for MCI vs. NCD classification with a balanced accuracy up to 0.71 (95% CI: 0.62-0.79). Multiclass classification yielded intermediate results, with balanced accuracy values up to 0.79 (CI: 0.74-0.84). Including 'possible' diagnoses appeared to reduce performance, particularly for the NCD versus MCI analysis. There were no statistical differences between ML models and a standard rule-based algorithm or between different ML models for objective 1 or objective 2. Key predictors of model performance included physician claims codes 331 (cerebral degeneration) and 290 (senile/presenile dementias). Conclusion: Machine learning algorithms applied to Canadian AHD can identify individuals who are clinically diagnosed with dementia with reasonable accuracy and is comparable to rule- based methods. However, reliably detecting MCI using AHD alone remains challenging. ML holds potential for dementia surveillance but requires richer AHD or novel features for identification of cognitive disorders.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.