Prediction of future dementia for MCI patients using a novel multimodal machine learning framework

Background Prediction of future dementia for patients with mild cognitive impairment (MCI) is a significant clinical goal, so that the identified cases can benefit from treatments. Currently, the clinical standard for diagnosing dementia utilizes multimodal data like cognitive tests, MRI and PET scans, and other biomarkers. However, efficiently and objectively analyzing these complex, diverse data can be difficult in the clinical setting, contributing to high rates of underdiagnosis or misdiagnosis of dementia at early stages. Machine learning (ML) offers a potentially more efficient and objective methodology to aid the prediction of future dementia for MCI patients from these multimodal data. Method e recently developed Ensemble Integration (EI), an ML framework to advance predictive modeling from multimodal data by leveraging complementarity among the data modalities (Li et al, Bioinformatics Advances, 2022). In the current work, we conducted the first assessment of EI’s ability to predict the future development of dementia among MCI patients using multimodal data from The Alzheimer’s Disease Prediction of Longitudinal Evolution (TADPOLE) challenge (Marinescu et al, Predictive Intelligence in Medicine, 2019) (Table 1). We developed an EI‐based predictive model of dementia from data of 672 MCI patients collected at their first visit (baseline), and rigorously evaluated this model and benchmark methods on baseline data from two separate test sets (Figure 1). Result For predicting the future development of dementia among MCI patients, the EI‐based model performed better on two test sets (AUROC = 0.77/0.78, sensitivity = 0.71/0.75, specificity = 0.74/0.75) than the more commonly used XGBoost (AUROC = 0.66/0.67, sensitivity = 0.59/0.63, specificity = 0.73/0.71) and deep learning (AUROC = 0.63/0.64, sensitivity = 0.78/0.62, specificity = 0.41/0.55) approaches. This model also suggested MRI‐derived measurements of the cortical thickness, surface area and volume of regions in the thalamus and parahippocampus to be predictive of this outcome (Table 2). Conclusion EI is an effective framework for predicting if an MCI patient will develop dementia in the future, as well as identifying neuroanatomical features that may be associated with this progression.