Development and Validation of a Machine Learning Classification Algorithm for Differentiating Frontotemporal Dementia
Seung Hyun Lee1, Wooseok Jung1, Mina Park2
1From the VUNO Inc. (S.H.L., W.J., H.C.), Seoul, Korea; UC Berkeley-UCSF Joint Graduate Program in Bioengineering (W.J.); Department of Radiology (M.P., H.S.O., B.J., S.J.A., S.H.S.), Neurology (H.C.C.H.L.), Gangnam Severance Hospital, Yonsei University College of Medicine, Seoul, Korea and Osstem Implant Co., Ltd., 3 (H.C.), Magokjungang 12-ro, Gangseo-gu, Seoul, Korea.
Background And Purpose:
Frontotemporal dementia (FTD) is characterized by frontal and anterior temporal lobe atrophy and progressive changes in behavior, executive function, or language, and its clinical and imaging presentations may overlap with Alzheimer's disease (AD), complicating differential diagnosis. This study aimed to develop and validate a classification algorithm using automated brain volumetry to differentiate FTD from AD and cognitively normal (CN) individuals, incorporating asymmetry indices, external validation in a real-world clinical setting, and a radiologist reader study to assess clinical utility.
Materials And Methods:
This study included 758 subjects (FTD = 142, AD = 216, CN = 400) from the NIFD and ADNI databases for training and internal validation, and 89 subjects (FTD = 18, AD = 24, CN = 47) from a tertiary hospital for external validation. Automated brain volumetry was obtained from 3D T1-weighted MRI using deep learning, and regional volumes were normalized to intracranial volume with age- and sex-adjusted z-scores. Volumetric features were grouped into frontal, medial, and lateral regions, with left-right and anteroposterior asymmetry indices. An XGBoost classifier was trained using these features and MMSE scores to distinguish FTD, AD, and CN. A reader study assessed diagnostic performance and interpretation time with and without algorithm assistance.
Results:
The classifier achieved 91.40% accuracy, 87.02% sensitivity, 95.09% specificity, and an AUC of 95.41% in the internal validation cohort, using a feature set that included frontal, medial, and lateral regional volumes, asymmetry metrics, and residual-based ICV normalization. External validation yielded 88.80% accuracy, 82.87% sensitivity, 94.41% specificity, and an AUC of 92.94%. FTD-specific sensitivity and specificity in the external validation cohort were 61.07% (95% CI, 38.45-84.01) and 97.13% (95% CI, 92.86-100.00), respectively. Among experienced readers, accuracy was maintained, with a significant reduction in reading time (p < 0.001) when assisted by the algorithm. For the underexperienced reader, accuracy improved without a significant change in interpre tation time (p = 0.786).
Conclusions:
The automated brain volumetry model demonstrated promising diagnostic accuracy for FTD and supported by asymmetry-based feature engineering, real-world external validation, and a radiologist reader study, may serve as a useful adjunct for differentiating FTD from AD and normal aging.
