Background and objectives: Distinguishing neurodegenerative diseases is a challenging task requiring neurologic expertise. Clinical decision support systems (CDSSs) powered by machine learning (ML) and artificial intelligence can assist with complex diagnostic tasks by augmenting user capabilities, but workflow integration poses many challenges. We propose that a modeling framework based on fluorodeoxyglucose PET (FDG-PET) imaging can address these challenges and form the basis of an effective CDSS for neurodegenerative disease.
Methods: This retrospective study focused on FDG-PET images in a discovery cohort drawn from 3 research studies plus routine clinical patients. When selecting research study participants, the inclusion criterion was the availability of an FDG-PET image from within 2.5 years of diagnosis with 1 of 9 specific neurodegenerative syndromes or designation as unimpaired. Participants from disease groups were recruited from the clinical patient population while unimpaired participants came primarily from a population study. The discovery cohort was used to develop a clinical decision support framework we call StateViewer, which applies a neighbor matching algorithm to detect the presence of 9 different neurodegenerative phenotypes. The ML performance of this framework was evaluated in the discovery cohort by nested cross-validation and externally validated in the Alzheimer's Disease Neuroimaging Initiative. Potential for clinical integration was demonstrated in a radiologic reader study focused on differentiating posterior cortical atrophy from Lewy body dementia.
Results: The discovery cohort contained 3,671 individuals with a mean age of 68 years and consisted of 49% reported female. Our model framework was able to detect the presence of 9 different neurodegenerative phenotypes with a sensitivity of 0.89 ± 0.03 and an area under the receiver operating characteristic curve of 0.93 ± 0.02. In the radiologic reader study, readers using our model were found to have 3.3 ± 1.1 times greater odds of making a correct diagnosis than readers using a current standard-of-care workflow.
Discussion: Our proposed framework provides strong classification performance with high interpretability, and it addresses many of the challenges that face clinical integration of ML-based decision support tools. One limitation of this study is a uniform discovery cohort that is not representative of other patient populations in some regards.