Predicting Alzheimer's Disease Progression from Sparse Multimodal Data by NeuralODE Models

Abstract

Alzheimer's disease shows significantly variable progressions between patients, making early diagnosis, disease monitoring, and care planning difficult. Existing data-driven Disease Progression Models try to tackle this issue, but they usually require sufficiently large datasets of specific diagnostic modalities, which are rarely available in clinical practice. Here, we introduce a new modeling framework capable of predicting individual disease trajectories from sparse, irregularly sampled, multi-modal clinical data. Our method uses (recurrent) Neural Ordinary Differential Equations to determine the current hidden state of a patient from sparse past exams and to forecast future disease progression, illustrating how biomarkers evolve over time. When applied to the ADNI clinical cohort, the model detected early signs of disease more accurately than common data-driven alternatives and effectively tracked changes in biomarker trajectories that align with established clinical knowledge. This provides a versatile tool for accurate diagnosis and monitoring of neurodegenerative diseases.

Keywords: Alzheimer’s disease; disease progression modeling; machine learning; neurodegenerative disorders; patient-specific trajectories; sparse and multimodal clinical data.

Figure 1.. Schematic representation of the NeuralODE-based DPM.

a, conceptual representation of available scattered data for three different patients, with each symbol representing a distinct category of clinical exams (Neuropsychological assessment, cerebrospinal fluid analysis, MRI, PET, other data). b, ENCODER architecture that compresses sparse time series into a dense latent state, encoding the patient’s health status. c, ODE-AE architecture that predicts exam results at each time point. I, U, and D stand for INIT-NN, UPDATE-NN, and DECODER-NN, respectively.

Figure 2.. Early diagnosis and feature reconstruction analyses of the model.

a, confusion matrix for the future diagnosis of currently healthy patients. b, mean absolute error for the reconstruction of several examinations, computed on normalized features. c, decomposition of mean absolute error into initial estimation error and evolution error. The list of acronyms used is detailed in Table 4.

Figure 3.. Average trajectories for healthy, MCI, and dementia patients.

The lines represent the means of the predicted time series, and the shaded areas represent the standard deviation of the predicted time series. a, typical trajectories of selected AD markers. b, Average trajectories of the latent state variables. Refer to Table 4 for the list of acronyms.

Figure 4.. Stream plot of the average evolution of the amyloid-PET, tau-PET, MMSE, and CDR scores.

The streamlines fill the space of values for which there are samples in the dataset, and the color represents the density of the samples. Refer to Table 4 for the list of acronyms.

Figure 5.. Average Jacobian matrices of several normalized AD markers reconstructed at the current time and after 5 years w.r.t. the normalized latent variables at the current time.

The features belonging to the same exam (e.g., the questions of the ADAS questionnaire) have been averaged together. Refer to Table 4 for the meaning of the acronyms.

Figure 6.. Prediction of several features for three patients:

healthy (left), MCI (center), and dementia (right). The line represents the mean of the predicted time series with the corresponding confidence band, the circles represent the input values, and the diamonds represent the future values. Refer to Table 4 for the meaning of the acronyms.

Figure 7.

Interpolation and forecast errors versus the number of patients in the training set.