L2C-FNN: Longitudinal to Cross-sectional FNN for Generalizable AD-dementia Progression Prediction
Poster No:
1390
Submission Type:
Abstract Submission
Authors:
Chen Zhang1, Lijun An1, Naren Wulan1, Csaba Orban1, Kim-Ngan Nguyen1, Pansheng Chen1, Christopher Chen1, Juan Helen Zhou1, B. T. Thomas Yeo1
Institutions:
1National University of Singapore, Singapore, Singapore
First Author:
Co-Author(s):
Introduction:
Alzheimer's disease dementia (AD-dementia) is a neurodegenerative disorder with a prolonged prodromal phase and limited therapeutic options post-dementia onset, emphasizing the importance of early detection for timely and effective intervention (Scheltens et al., 2016). Hence, predicting longitudinal disease progression of individuals is of substantial interest (Ghazi et al., 2019; Nguyen et al., 2020). However, the absence of cross-cohort assessments in previous studies have raised concerns about clinical applicability (Wang et al., 2022) due to the cohort disparities. In this study, we introduce the Longitudinal to Cross-sectional Feedforward Neural Network (L2C-FNN), a robust model designed to mitigate cohort differences and demonstrate its superior generalizability against strong machine learning baseline models across three separate unseen cohorts.
Methods:
L2C-FNN and baseline models underwent training on ADNI (N=2421; Jack et al., 2010) followed by evaluation of generalizability on external test cohorts: AIBL (N=862; Ellis et al., 2010) from Australia, MACC (N=700; Hilal et al., 2020) from Singapore, and OASIS (N=1378; LaMontagne et al., 2019) from North America. ADNI participants were randomly divided into training, validation, and test sets (ratio of 18:1:1) for model fitting, hyperparameter tuning and within-cohort evaluation. The trained models were adapted to AIBL, MACC, and OASIS for cross-cohort evaluation, with 20 repetitions to ensure result stability (Figure 1A). Care was taken to ensure non-overlapping test sets, covering the entirety of the ADNI cohort across the 20 data splits.
Utilizing multimodal inputs (e.g., cognitive state measurements, cortical and/or subcortical ROI volumes) from the first 50% of timepoints of each participant, we predicted clinical diagnosis, ventricular volume, and cognitive state for the second 50% of timepoints, projecting up to 10 years into the future. All continuous variables (e.g., ROI volumes) underwent normalization through GaussRank transformation, a special form of quantile normalization (Zhao et al., 2020), with a Gaussian reference distribution.
L2C-FNN (Figure 1B) is a deep feedforward neural network featuring a specialized longitudinal-to-cross-sectional format transformation, which involves computing summary statistics such as the rate of change, maximum, and minimum of each input modality from historical timeseries data. The L2C transformation offers an advantage by eliminating the reliance on error-prone recursive techniques like RNN commonly used in disease progression modeling (Fan et al., 2019).
Baseline approaches included Frog, an XGBoost-based model, and MinimalRNN (Nguyen et al., 2020) an RNN-based model, which were 1st and 2nd place winners in the TADPOLE international challenge for longitudinal AD-dementia progression prediction (Marinescu et al., 2021).
Utilizing multimodal inputs (e.g., cognitive state measurements, cortical and/or subcortical ROI volumes) from the first 50% of timepoints of each participant, we predicted clinical diagnosis, ventricular volume, and cognitive state for the second 50% of timepoints, projecting up to 10 years into the future. All continuous variables (e.g., ROI volumes) underwent normalization through GaussRank transformation, a special form of quantile normalization (Zhao et al., 2020), with a Gaussian reference distribution.
L2C-FNN (Figure 1B) is a deep feedforward neural network featuring a specialized longitudinal-to-cross-sectional format transformation, which involves computing summary statistics such as the rate of change, maximum, and minimum of each input modality from historical timeseries data. The L2C transformation offers an advantage by eliminating the reliance on error-prone recursive techniques like RNN commonly used in disease progression modeling (Fan et al., 2019).
Baseline approaches included Frog, an XGBoost-based model, and MinimalRNN (Nguyen et al., 2020) an RNN-based model, which were 1st and 2nd place winners in the TADPOLE international challenge for longitudinal AD-dementia progression prediction (Marinescu et al., 2021).
Results:
Figure 2A demonstrates the comparable performance of L2C-FNN with strong baseline methods (Frog and MinimalRNN) for within-cohort (ADNI) clinical diagnosis, cognitive state, and ventricular volume prediction, Notably, L2C-FNN clinical diagnosis and MMSE prediction outperformed all baselines numerically.
Figure 2B shows cross-cohort evaluation in three external cohorts (AIBL, MACC, and OASIS), highlighting the superior performance of L2C-FNN over all baseline models, underscoring its robust generalizability. Particularly noteworthy is L2C-FNN's consistent achievement of significantly lower MMSE prediction errors across all test cohorts compared to the baseline methods.
Figure 2B shows cross-cohort evaluation in three external cohorts (AIBL, MACC, and OASIS), highlighting the superior performance of L2C-FNN over all baseline models, underscoring its robust generalizability. Particularly noteworthy is L2C-FNN's consistent achievement of significantly lower MMSE prediction errors across all test cohorts compared to the baseline methods.
Conclusions:
Our findings demonstrate the superiority of the L2C-FNN model over baseline algorithms in longitudinal AD-dementia progression prediction when trained and tested on ADNI dataset. Crucially, this strong performance extended to previously unseen cohorts with significantly diverse populations from the training set, including AIBL, MACC, and OASIS, as confirmed by cross-cohort evaluation, emphasizing the model's superior generalizability.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 2
Lifespan Development:
Aging
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Methods Development
Multivariate Approaches
Keywords:
Aging
Data analysis
Degenerative Disease
Machine Learning
Memory
Multivariate
Statistical Methods
STRUCTURAL MRI
Other - Longitudinal Forecasting; Generalizable Models; Deep Learning
1|2Indicates the priority used for review
Provide references using author date format
[2] Fan, C. (2019), 'Assessment of deep recurrent neural network-based strategies for short-term building energy predictions', Applied Energy, 236, 700–710.
[3] Hilal, S. (2020), ‘Cortical cerebral microinfarcts predict cognitive decline in memory clinic patients’, Journal of Cerebral Blood Flow and Metabolism: Official Journal of the International Society of Cerebral Blood Flow and Metabolism, 40(1), 44–53.
[4] Jack, C. R. (2010), ‘Update on the magnetic resonance imaging core of the Alzheimer’s disease neuroimaging initiative’, Alzheimer’s & Dementia: The Journal of the Alzheimer’s Association, 6(3), 212–220.
[5] LaMontagne, P. J. (2019), ‘OASIS-3: Longitudinal Neuroimaging, Clinical, and Cognitive Dataset for Normal Aging and Alzheimer Disease’, medRxiv, 2019.12.13.19014902.
[6] Marinescu, R. V. (2021), ‘The Alzheimer’s Disease Prediction Of Longitudinal Evolution (TADPOLE) Challenge: Results after 1 Year Follow-up' (arXiv:2002.03419). arXiv.
[7] Mehdipour, Ghazi. (2019), ‘Training recurrent neural networks robust to incomplete data: Application to Alzheimer’s disease progression modeling', Medical Image Analysis, 53, 39–46.
[8] Nguyen, M. (2020), ‘Predicting Alzheimer’s disease progression using deep recurrent neural networks’, NeuroImage, 222, 117203.
[9] Scheltens, P. (2016), ‘Alzheimer’s disease’, The Lancet, 388(10043), 505–517.
[10] Wang, C. (2022), ‘A high-generalizability machine learning framework for predicting the progression of Alzheimer’s disease using limited data’, Npj Digital Medicine, 5(1), Article 1.
[11] Zhao, Y. (2020), ‘How to do quantile normalization correctly for gene expression data analyses’, Scientific Reports, 10(1), Article 1