Source-linked AI summary
Deep learning for comprehensive forecasting of Alzheimer's Disease progression
Charles K. Fisher, Aaron M. Smith, Jonathan R. Walsh, the Coalition Against Major Diseases
TL;DR
The work addresses limitations in applying machine learning to broader patient-level forecasting. It uses unsupervised deep learning to simulate entire patient trajectories, matching specific models on individual prediction tasks while capturing variable correlations.
Problem
Broader applications of machine learning remain limited, motivating higher-confidence data-driven patient-level forecasts.
Method
A latent variable model uses unsupervised deep learning to create stochastic simulations of entire patient trajectories.
Results
The approach achieves the same performance as specific models on individual prediction tasks while accurately capturing correlations between variables; word-recall measures identify patients who progress more rapidly.
Takeaways & Limitations
Patient-level simulations are technologically feasible as a proof of concept and may support future research.
Takeaways & Limitations
The authors describe the dataset and methodology as limited and identify cognitive-exam performance that cannot be predicted from baseline measures.
Abstract
from arXiv · showhide
Most approaches to machine learning from electronic health data can only predict a single endpoint. Here, we present an alternative that uses unsupervised deep learning to simulate detailed patient trajectories. We use data comprising 18-month trajectories of 44 clinical variables from 1908 patients with Mild Cognitive Impairment or Alzheimer's Disease to train a model for personalized forecasting of disease progression. We simulate synthetic patient data including the evolution of each sub-component of cognitive exams, laboratory tests, and their associations with baseline clinical characteristics, generating both predictions and their confidence intervals. Our unsupervised model predicts changes in total ADAS-Cog scores with the same accuracy as specifically trained supervised models and identifies sub-components associated with word recall as predictive of progression. The ability to simultaneously simulate dozens of patient characteristics is a crucial step towards personalized medicine for Alzheimer's Disease.
I. INTRODUCTION
Existing clinical machine-learning approaches often target a single endpoint, while heterogeneous Alzheimer’s disease progression requires detailed, patient-level forecasting across many characteristics. This work addresses that gap with unsupervised deep learning that simulates entire patient profiles and their evolution.
- Patients with the same disease can show different symptoms, progression rates, and responses to therapy, motivating precision medicine approaches.
- Most machine-learning applications for electronic health data use supervised learning to predict specific endpoints.
- The proposed direction is to use deep learning to simulate detailed patient progression and support broader applications of machine learning.
- Clinical health data are multimodal, relatively small, and frequently incomplete, creating challenges for current deep-learning methods.
- Prior Alzheimer’s disease forecasting models generally predict a single endpoint, such as change in the ADAS-Cog score.
- The study simulates complete patient profiles, including cognitive-score subcomponents, laboratory tests, and associations with baseline clinical characteristics.
A. Data Processing
The study extracted 18-month trajectories for 1908 patients with MCI or AD, representing 44 variables at 3-month intervals. Data processing standardized heterogeneous clinical records, selected sufficiently observed variables, and separated training, validation, and testing cohorts.
- Time-dependent measurements were aggregated into 90-day windows, while static variables were reduced to single estimates and units were harmonized.
- Categorical variables were one-hot encoded, continuous variables were transformed and standardized, and values were converted back before analysis.
- The analysis selected 44 variables observed in a reasonably large fraction of patients because heavily missing covariates could impair performance.
- The dataset contains 18-month trajectories from 1908 patients with MCI or AD across 44 cognitive, laboratory, clinical, and background variables.
- Patient trajectories represent all 44 variables at seven 3-month time points from baseline through 18 months.
B. Machine Learning
The model is a generative Conditional Restricted Boltzmann Machine that learns distributions over multimodal patient variables and iteratively generates longitudinal trajectories. It supports missing-data imputation and both baseline-conditioned and entirely synthetic patient simulations.
- The CRBM models time dependence between neighboring time points, while different visible-layer units represent multimodal observations.
- Missing observations are imputed by sampling their conditional distribution given observed values.
- Generative modeling draws baseline patient states and repeatedly samples conditional future states to produce complete trajectories.
- The study uses a Conditional Restricted Boltzmann Machine to model patient progression as a latent-variable generative process.
- The model uses latent variables to represent correlations among covariates and trains one model across neighboring time points.
- It generates either trajectories initialized from real patient baselines or entirely synthetic patients by simulating baseline data and subsequent time points.
A. General model performance
The CRBM reproduced marginal distributions and both equal-time and lagged correlations across the clinical variables, indicating strong generative performance. It also provided a unified framework for forecasting all time-dependent variables.
- The synthetic trajectories accurately captured marginal distributions of cognitive scores, laboratory tests, and clinical data from real patients.
- Synthetic patient trajectories modeled observed variable correlations well, including strong within-category relationships and weaker cross-category relationships.
- The model reproduced equal-time and lagged autocorrelations, including lags longer than 3 months, despite directly connecting only neighboring time points.
- Figure 2 evaluates generative performance by comparing observed and predicted correlations and autocorrelations across time lags.
- A single CRBM predicts all variables and time points, with forecast errors compared against a random forest trained for each variable and time point.
B. Simulating conditional patient trajectories
The CRBM simulates complete, correlated patient trajectories and supports conditional forecasts for observed or unobserved characteristics. Its accuracy approaches specialized random-forest models while preserving trajectory-level correlations and uncertainty.
- Modeling approach: A single CRBM predicts all variables and time points, replacing 210 separate random-forest models for variable–time-point combinations.The random-forest comparison used 35 time-dependent variables across six future time points.
- Predictive accuracy: The CRBM’s accuracy is close to specialized random-forest models for each variable and time point, generally outperforming them on ADAS-Cog components.
- Correlated trajectories: Unlike a factorized model, the CRBM maintains correlations among covariates while achieving equivalent accuracy on individual prediction problems.The factorized model cannot generate realistic trajectories with correlated covariates.
- Uses of simulation: The framework can simulate entire patient profiles, explore alternative endpoints and subgroups, and aggregate individual risk estimates to population-level risks.
- Uncertainty: The model provides forecasts and confidence intervals for multiple patient characteristics, supporting estimates of expected progression and uncertainty.
C. Forecasting and interpreting disease progression
The model forecasts disease progression over time and helps interpret heterogeneity in progression rates. It matches supervised approaches for ADAS-Cog forecasting and links faster progression to recall-related cognitive measures.
- Population progression: The model reproduces the observed increase in mean ADAS-Cog scores and widening right tail over time, indicating that a subset of patients drives much of the population trend.
- Forecasting: CRBM predictions remain accurate through at least 18 months for changes in ADAS-Cog score from baseline.
- Model comparison: The CRBM is the best-performing model for the ADAS-Cog forecasting task, although model accuracies converge at study durations of 15 months or longer.
- Model comparison: The CRBM performs strongly despite training only on 3-month-ahead simulations and not directly on aggregate ADAS-Cog scores.
- Progression correlates: Most baseline variables are not associated with progression, whereas recall-related tests and word recognition show strong associations.
- Progression correlates: Poor ADAS delayed word recall predicts faster progression even after controlling for total ADAS-Cog score.ADAS delayed word recall is an additional disease-severity probe and is outside the 11-component ADAS-Cog score.
IV. DISCUSSION
The model simulates stochastic, individualized Alzheimer’s disease trajectories across many clinical variables while matching specialized models on individual prediction tasks. It also captures variable correlations, provides uncertainty estimates, and identifies word recall as predictive of faster progression.
- Model capabilities: Unsupervised deep learning creates stochastic simulations of entire patient trajectories while matching specific models on individual prediction tasks.The approach also captures correlations between variables.
- Model capabilities: Generative models provide forecasts, uncertainty estimates, and simultaneous assessment of different risks.These capabilities address the need for individual-level forecasts that include expected outcomes and their uncertainty.
- Model capabilities: The model integrates multimodal continuous, discrete, time-dependent, and static variables within one framework.It can also handle missing training observations through automated imputation, reducing preprocessing requirements for synthetic clinical data.
- Model assumptions: A single time-lagged connection was sufficient to explain temporal correlations in Alzheimer’s disease, although more complex diseases may require additional connections.The required temporal model complexity may therefore depend on disease characteristics.
- Clinical findings: ADAS-Cog and MMSE components are only weakly correlated with other clinical variables.This supports examining individual cognitive components rather than relying exclusively on aggregate cognitive scores.
- Limitations: Observed stochasticity may reflect heterogeneity in cognitive-exam performance that baseline measurements cannot predict.This possibility limits confidence in individual-level forecasts.
- Clinical findings: Poor performance on baseline word recall tests predicts faster progression even after controlling for the ADAS-Cog score.The finding identifies an individual cognitive component associated with progression beyond the total cognitive score.
V. CONCLUSIONS
The work demonstrates a proof-of-concept for patient-level simulations of Alzheimer’s disease progression, while identifying data and methodological limitations that constrain its current scope. The approach could support broader clinical-data simulation and personalized risk forecasting with suitable tools and data.
- Conclusions: Patient-level simulations are technologically feasible for forecasting Alzheimer’s disease progression.The authors describe this as a proof-of-concept rather than a completed clinical system.
- Conclusions: The model currently represents 44 variables commonly measured in Alzheimer’s disease clinical trials.Some potentially informative covariates were excluded because they were not measured in most patients.
- Conclusions: The dataset lacks neuroimaging studies and amyloid-β level measurements.The authors identify incorporating additional data as an important next step, especially as surrogate biomarkers become standard in clinical trials.
- Conclusions: The simulation approach can be extended to other diseases and used to generate synthetic datasets or simulated clinical trials.The authors also describe possible uses as synthetic control arms and for forecasting individual risks to support treatment choices.
- Conclusions: Limited availability of high-quality longitudinal health datasets and limited ability of current methods to produce useful insights slow progress toward these applications.The authors qualify that progress is feasible with the right tools and data.
VII. SUPPORTING INFORMATION
The supporting information describes a repeatable pipeline for converting structured clinical-trial data into patient-level machine-learning datasets. It also explains how conditional restricted Boltzmann machines model multimodal longitudinal data, missingness, and uncertainty.
- Data processing: The pipeline converts CAMD clinical-trial data into machine-learning-ready datasets through raw and processed SQL databases.Data are ingested after abbreviation translation and simple type inference, then transformed and stored for dataset construction.
- Study dataset: The dataset uses 44 variables measured at 3-month intervals from baseline through 18 months.The 3-month interval was chosen because most patients had data at each time point, whereas shorter intervals produced missing time points.
- Modeling framework: Boltzmann machines support multimodal clinical variables, conditional and generative sampling, missing-data imputation, and uncertainty estimation.Stochastic sampling produces both predictions and uncertainty, while the learned joint distribution can impute missing values.
- Modeling framework: A CRBM models adjacent time points jointly, allowing patient trajectories and future clinical data to be sampled conditionally from baseline information.The model uses visible units for multiple time points and static variables, with k = 1 so two adjacent time points are learned simultaneously.
C. Details of Training
The training procedure fits a CRBM to adjacent patient-time-point pairs using a combined likelihood and adversarial objective, then evaluates its stochastic forecasts against supervised models. Training eventually learns the observed progression pattern, although early training regresses toward the mean for ADAS-Cog outliers.
- Details of Training: The CRBM trains on adjacent pairs of patient time points, concatenating consecutive temporal data with static variables.Each patient contributes six adjacent pairs from seven available time points, and samples are shuffled into minibatches.
- Details of Training: C = −γL −(1 −γ)A combines log-likelihood and adversarial objectives, with γ weighting their relative contributions.The adversarial component uses a random forest classifier.
- Training dynamics: Within 100 epochs, monitored divergence metrics approach their final values, but the model initially performs worse for patients with extreme ADAS-Cog scores.The early model exhibits regression toward the mean for ADAS-Cog outliers.
- Training dynamics: Continued training allows the CRBM to learn that higher-scoring individuals progress more rapidly.This replaces the early pattern in which low-scoring patients progressed rapidly and high-scoring patients improved.
- Evaluation: Supervised models require separate models for each readout time, whereas the same CRBM is used across all readout times.All algorithms are evaluated on the same test set using RMS error.
- Evaluation: The CRBM predicts ADAS-Cog change by repeatedly simulating trajectories from baseline and averaging each patient’s distribution of simulated score changes.The distribution also supplies a probabilistic representation of the CRBM’s predictions and their standard-error evaluation.