Friday, October 24, 2025

Predicting Alzheimer’s Disease Risk: A Comorbidity-Based Survival Analysis Using Machine Learning

Share

Exploring Participant Demographics and Clinical Features in Alzheimer’s Disease Research

The Alzheimer’s Disease Neuroimaging Initiative (ADNI) and the Australian Imaging, Biomarker & Lifestyle Flagship Study of Ageing (AIBL) offer invaluable datasets for understanding the progression of Alzheimer’s disease. This article delves into the demographic and clinical characteristics of participants from ADNI and AIBL, focusing on how these features stratify within censored and transitioned groups.

Participant Demographics and Clinical Features

Demographics of ADNI and AIBL

In the ADNI dataset, we find a total of 389 participants categorized as "censored" (those who have not transitioned to mild cognitive impairment, or MCI) and 105 who have "transitioned" (those who have progressed to MCI). In contrast, AIBL presents data on 290 censored individuals and 30 transitioned individuals. Participants within ADNI demonstrate a statistically significant age difference: the censored group has a mean age of 73.84 years (SD = 5.84), while the transitioned group averages 75.80 years (SD = 5.65), a difference significant at (p = 0.002). Although the gender distribution leans slightly towards more males in the transitioned group (56.2%) compared to the censored group (47.0%), this is not statistically significant ((p = 0.120)).

Cognitive Scores and Comorbidities

Cognitive assessments reveal notable disparities between the two groups in several metrics. For instance, the Clinical Dementia Rating Scale Sum of Boxes (CDRSB), Alzheimer’s Disease Assessment Scale (ADAS), Rey Auditory Verbal Learning Test (RAVLT), and Functional Activities Questionnaire (FAQ) each show significant differences that underscore the cognitive decline associated with transition to MCI.

In terms of health profiles, certain comorbidities reveal acute differences. For example, the prevalence of Endocrine Metabolic conditions stands at 46.3% in censored individuals versus 57.1% in transitioned individuals ((p = 0.062)), while the percentage of Renal Genitourinary conditions is significantly higher in the transitioned group (42.2% vs. 57.1%, (p = 0.009)). Comparing these findings to AIBL, the demographic trends show that AIBL participants are slightly younger on average (72.32 years) but maintain similar education levels and racial distribution.

Variations in Health Profiles

The health profiles of AIBL participants differ significantly. Endocrine and metabolic conditions exist at much lower prevalence rates in AIBL (below 17% for both transitioned and censored groups), contrasting sharply with ADNI. Such differences highlight potential variations in lifestyle, healthcare access, or genetic predispositions affecting these populations.

Performance of Predictive Models

Model Evaluation in ADNI

To enhance our understanding of early-stage Alzheimer’s disease progression, a range of machine learning and deep learning models have been assessed against the ADNI dataset. Models such as Cox proportional hazards (CoxPH), random survival forest (RSF), and DeepSurv, among others, were evaluated across various feature sets combining demographic data, cognitive scores, and comorbidities.

The performance results showcased in a heatmap (Fig. 1) indicate commendable predictive accuracy. Notably, both the fast random survival forest and DeepSurv models achieved a C-index of 0.84 when using a comprehensive feature set (FS1). The comparative analysis reveals that these top models significantly surpass prior studies on survival analysis of CN (Cognitive Normal) to MCI conversions within the same dataset, where previous C-indices hovered around 0.66 to 0.68.

Statistical Significance of Model Performance

Model performance evaluation employed techniques such as the Kruskal–Wallis test, which identified overarching differences among model predictions. Post-hoc testing identified that the fast random survival forest significantly outperformed many other models, providing both statistical robustness and practical relevance in clinical decision-making contexts.

Factors Influencing Predictions

Key Predictive Features

The two standout models, fast random survival forest and DeepSurv, indicate that certain features play a critical role in survival predictions. With features like age, various cognitive scores (ADAS13, RAVLT), and the presence of comorbidities prominently influencing outcomes, the analysis highlights age as a major risk factor, corroborating established research findings on Alzheimer’s disease.

Visualizing Feature Impact

Partial Dependence Plots (PDPs) further elucidate the relationships between key predictors and survival outcomes. For example, various plots convey that as age increases, the associated survival probability diminishes, highlighting an urgent need for early interventions in older populations.

Kaplan-Meier Survival Curves

The analysis extends to Kaplan-Meier curves which stratify survival based on critical variables such as age and comorbidities. For instance, individuals with renal comorbidities demonstrated significantly poorer survival outcomes compared to those without (log-rank (p = 0.04)). Age also emerged as a critical predictor, with those older than 70 years exhibiting lower survival probabilities (log-rank (p = 0.003)).

External Validation

To ensure the generalizability of the fast random survival forest model, an external validation step was performed utilizing the AIBL dataset. The model displayed a commendable C-index of 0.73 when applied directly to AIBL, further enhancing its reliability. Rerunning the model on only shared features yielded a C-index of 0.75, indicating stable performance across datasets. Notably, employing multiple imputation improved predictive outcomes, suggesting that handling missing data effectively is critical in clinical research contexts.

The ongoing exploration within the ADNI and AIBL datasets exemplifies the potential of machine learning techniques in predicting Alzheimer’s disease progression, helping to unveil critical insights that could lead to improved patient outcomes. By merging robust statistical methodologies and practical clinical applications, this research promises to advance understanding in this critical area of geriatric health.

Read more

Related updates