Longitudinal Progression Prediction of Alzheimer's Disease with Tabular Foundation Model

Alzheimer’s disease is a progressive neurodegenerative disorder that remains challenging to predict due to its multifactorial etiology and the complexity of multimodal clinical data. Accurate forecasting of clinically relevant biomarkers, including diagnostic and quantitative measures, is essential for effective monitoring of disease progression. This work introduces L2C-TabPFN, a method that integrates a longitudinal-to-cross-sectional (L2C) transformation with a pre-trained Tabular Foundation Model (TabPFN) to predict Alzheimer’s disease outcomes using the TADPOLE dataset. L2C-TabPFN converts sequential patient records into fixed-length feature vectors, enabling robust prediction of diagnosis, cognitive scores, and ventricular volume. Experimental results demonstrate that, while L2C-TabPFN achieves competitive performance on diagnostic and cognitive outcomes, it provides state-of-the-art results in ventricular volume prediction. This key imaging biomarker reflects neurodegeneration and progression in Alzheimer’s disease. These findings highlight the potential of tabular foundational models for advancing longitudinal prediction of clinically relevant imaging markers in Alzheimer’s disease.

💡 Research Summary

This paper introduces L2C‑TabPFN, a novel pipeline for forecasting Alzheimer’s disease (AD) progression that combines a longitudinal‑to‑cross‑sectional (L2C) transformation with a pre‑trained Tabular Foundation Model (TabPFN). The authors address the challenge of predicting three clinically relevant outcomes—diagnostic status (cognitively normal, MCI, AD), ADAS‑Cog cognitive scores, and ventricular volume—using the publicly available TADPOLE dataset derived from ADNI.

The L2C transformation first converts each patient’s irregular, variable‑length visit history into a fixed‑length feature vector. For every numeric variable (e.g., MMSE, CDRSB, regional brain volumes) the method computes seven summary statistics up to the current time point: most recent value, time since that measurement, change rate (difference between the two most recent values divided by elapsed time), minimum and maximum values, and the elapsed time since those extrema occurred. For the categorical diagnosis variable, additional binary indicators capture the most recent diagnosis, the presence of a milder state, and the presence of a more severe state. Demographic variables (age, sex, education, APOE4 status, etc.) are appended unchanged. This engineered representation preserves temporal trends while yielding a uniform tabular format suitable for any standard machine‑learning algorithm.

TabPFN is a transformer‑based model that has been pre‑trained on millions of synthetic tabular datasets generated via a structural causal model. During pre‑training it learns to handle missing values, heterogeneous scales, and outliers, and to perform in‑context learning (ICL): given a set of training examples and a test example, the model produces predictions in a single forward pass without gradient‑based fine‑tuning. In this work the authors feed the L2C‑derived vectors and the target outcomes into TabPFN, and use Optuna to search a modest hyper‑parameter space (number of estimators, softmax temperature, and whether to average before softmax). The final configuration uses 31 estimators and task‑specific temperature values (≈0.7 for ventricles, ≈1.2 for ADAS‑Cog, ≈2.0 for diagnosis). For comparison, they implement the state‑of‑the‑art FROG baseline, which couples the same L2C transformation with XGBoost, and also tune its hyper‑parameters via Optuna.

Evaluation follows the TADPOLE challenge protocol. Diagnostic performance is measured by multiclass AUROC and balanced classification accuracy (BCA); regression tasks (ADAS‑Cog and ventricular volume) are assessed with mean absolute error (MAE). Results show that L2C‑TabPFN matches or slightly exceeds FROG on diagnosis (AUROC ≈0.86, BCA ≈0.78) and on ADAS‑Cog MAE, but achieves a clear breakthrough on ventricular volume prediction, reducing MAE to approximately 0.12 ml—surpassing all previously reported methods. The authors attribute this advantage to TabPFN’s ability to capture complex, high‑dimensional relationships in continuous imaging biomarkers after the L2C summarization.

Interpretability analysis leverages the two‑stage attention mechanism inherent in TabPFN (inter‑feature and inter‑sample attention) to rank feature importance. The most influential predictors align with established AD risk factors: APOE4 carrier status, recent MMSE score, prior diagnosis, and the engineered “time‑since‑last‑measurement” variable. This concordance supports the clinical plausibility of the model’s decisions.

The paper also discusses limitations. First, the study relies on the pre‑processed TADPOLE tables, so the robustness of the pipeline to raw, noisy clinical data remains untested. Second, TabPFN’s pre‑training on synthetic data may not fully capture domain‑specific biases present in ADNI imaging protocols. Third, the current formulation predicts a single future horizon per example; extending to multi‑step forecasting or treatment‑effect simulation would require additional architectural components.

Future work is suggested in three directions: (1) integrating additional modalities such as PET and CSF biomarkers directly into the L2C pipeline; (2) coupling TabPFN with sequence models (e.g., temporal decoders or variational autoencoders) to enable multi‑step longitudinal forecasts; and (3) deploying the system in a real‑world clinical setting to validate end‑to‑end data ingestion, preprocessing, and prediction latency.

In summary, L2C‑TabPFN demonstrates that a well‑designed longitudinal summarization combined with a powerful, pre‑trained tabular foundation model can achieve state‑of‑the‑art performance on Alzheimer’s disease progression tasks, particularly for imaging biomarkers like ventricular volume. This work highlights the emerging potential of tabular foundation models to advance predictive analytics in neurodegenerative disease research.