Thi Nga Mai 1, Sungwoo Goo 1, Hyojin Cho 1, Mai Linh Hoang 1, Soyoung Lee 1, Hwi-yeol Yun 1,2,3, Jung-woo Chae 1,2,3
1 Chungnam National University (Daejeon, Republic of Korea), 2 Department of Bio-AI convergence, Chungnam National University (Daejeon, Republic of Korea), 3 Senior Health Convergence Research Center, Chungnam National University (Daejeon, Republic of Korea)
Objectives: The primary objective of this study was to develop an Explainable Deep Survival Analysis framework that predicts the progression of cognitive decline. Mild cognitive impairment (MCI) represents a transitional stage between healthy aging and dementia[1]. Focusing on the transitions from a Normal state to Mild Cognitive Impairment (MCI) and to Dementia (Dem), this study aims to use a Deep Weibull model to learn individual-specific scale parameters from patient features to ensure monotonic, age-related increases in hazard. By integrating explainable AI through SHAP (SHapley Additive exPlanations), the model identified key cognitive features driving disease progression. This enhances interpretability and provides personalized clinical insights into the domains that most strongly influence the timing of disease onset.
Methods: The study included 936 older adults who underwent clinical Dementia evaluation and completed both the K-MMSE and SNSB-II cognitive tests, with MCI patients defined as those who had a total K-MMSE score<24[2]. Data preprocessing involved labeling patients into three ordinal stages (Normal, MCI, and Dementia) and normalizing features.
The core architecture is a Multi-Layer Perceptron (MLP)[3] with Softplus activation to ensure all predicted parameters remain positive. The model estimates two individual-specific scale parameters (λMCI and λDem), while a shared shape parameter k>1 represents the increasing hazard rate associated with the aging process. We implemented a crucial medical monotonicity constraint (ScaleDem=ScaleMCI+Diff) to guarantee that predicted Dementia onset follows MCI. A custom parametric negative log-likelihood based on the Weibull CDF was used, allowing for effective handling of interval-censored data where the testing only reveals the patient’s state at that time.
For interpretability, SHapley Additive exPlanations[4] was used to attribute feature importance across cognitive domains and show how each feature affects the model’s output throughout the dataset. Performance underwent a thorough evaluation using time-dependent AUC and Brier Scores, with Bootstrap resampling employed to generate 95% Confidence Intervals.
Results: In the initial model using 13 cognitive domains, the AUC values for MCI and dementia were 0.83 and 0.90, with corresponding Brier scores of 0.1149 and 0.1153. Feature importance analysis identified an optimal subset of 8 domains, including Years of Education, three domains from SNSB-II, and four from K-MMSE, yielding higher AUC values for MCI and Dementia of 0.94 and 0.91, and lower Brier Scores of 0.0488 and 0.1164, respectively. Age-dependent analysis revealed that the Dementia model maintained higher stability (AUC 0.82–0.88) compared to the MCI model (AUC 0.70–0.75). Furthermore, the MCI model exhibited a significantly lower and more stable Brier Score (0.08) compared to the Dementia model (0.47), indicating high calibration for MCI probability predictions. SHAP Beeswarm plots were clinically plausible: Higher scores in Memory, Language, and Executive Function were associated with protective effects and reduced risk of cognitive decline. Higher years of education showed a negative SHAP effect, consistent with the neural reserve hypothesis. Survival simulations indicated that the risk of cognitive impairment increases significantly and follows a nearly linear trajectory between ages 70 and 85.
Conclusions: This research demonstrates that combining deep learning with parametric survival analysis provides a robust and interpretable method for predicting cognitive decline. The Deep Weibull model effectively captures the continuous nature of neurodegeneration while adhering to logical medical progression from Normal to Dementia states. By incorporating SHAP, the framework provides transparency, allowing clinicians to identify personalized risk factors and specific cognitive deficits that speed up disease onset. The results suggest that discrimination power is high for Dementia, and provide probability estimates for MCI that are highly calibrated, offering a comprehensive tool for early intervention and personalized patient management in aging populations.
N.Mai, S.Goo, H.Cho contributed equally to this work; corresponding authors: S. Lee, H.-y. Yun , J.-w. Chae
References:
[1] Anderson ND. State of the science on mild cognitive impairment (MCI). CNS Spectrums. 2019;24(1):78-87. doi:10.1017/S1092852918001347
[2] Folstein MF, Folstein SE, McHugh PR. “Mini-mental state”. A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res. (1975) 12:189–98.
[3] Popescu, Marius-Constantin, et al. “Multilayer perceptron and neural networks.” WSEAS transactions on circuits and systems 8.7 (2009): 579-588.
[4] Ponce-Bobadilla AV, Schmitt V, Maier CS, Mensing S, Stodtmann S. Practical guide to SHAP analysis: Explaining supervised machine learning model predictions in drug development. Clin Transl Sci. 2024 Nov;17(11):e70056. doi: 10.1111/cts.70056. PMID: 39463176; PMCID: PMC11513550.
Reference: PAGE 34 (2026) Abstr 12093 [www.page-meeting.org/?abstract=12093]
Poster: Methodology – AI/Machine Learning