L2C-TabPFN: Alzheimer’s Forecasting & Causal Probe

• The paper introduces L2C-TabPFN, which transforms variable-length clinical trajectories into fixed-length vectors for precise Alzheimer’s outcome prediction.
• It leverages a transformer-based Tabular Foundation Model to predict diagnosis, ADAS-Cog, and ventricular volume, achieving state-of-the-art results particularly in ventricular forecasting.
• A distinct variant uses frozen embeddings with a learnable decoder to recover causal structure, illustrating the model’s dual role in clinical forecasting and causal discovery.

L2C-TabPFN most commonly denotes a forecasting pipeline that combines a longitudinal-to-cross-sectional (L2C) transformation with a pre-trained Tabular Foundation Model (TabPFN) for Alzheimer’s disease progression prediction on TADPOLE/ADNI-derived data. In this formulation, irregular, multimodal, variable-length patient trajectories are converted into fixed-length tabular snapshots, after which TabPFN is used to predict future diagnosis, cognitive scores, and ventricular volume. The term is context dependent in the recent literature, however, because a separate causal-discovery paper also uses “L2C-TabPFN” for an adapter that decodes causal structure from frozen TabPFNv2 embeddings rather than for longitudinal clinical forecasting (Ding et al., 25 Aug 2025, Swelam et al., 10 Nov 2025).

1. Terminology and conceptual scope

In the Alzheimer’s disease forecasting setting, “L2C” refers to a longitudinal-to-cross-sectional transformation. The purpose of the transformation is to make longitudinal clinical records compatible with a tabular foundation model that expects fixed-length vectors rather than variable-length sequences. The resulting system predicts three outcomes from TADPOLE data: diagnosis (DX), ADAS-Cog, and ventricular volume normalized by intracranial volume (Ding et al., 25 Aug 2025).

A distinct usage appears in causal discovery, where L2C-TabPFN is described as an adapter on top of TabPFNv2 that probes whether pretrained hidden states encode recoverable causal structure. There, frozen embeddings are passed to a learnable decoder driven by causal tokens and converted into probabilistic adjacency matrices for directed edge prediction under DAG constraints (Swelam et al., 10 Nov 2025).

This naming overlap is consequential. In one line of work, L2C-TabPFN is a clinical time-series-to-table forecasting pipeline; in the other, it is a frozen-backbone causal-discovery probe. The two systems share TabPFN as a substrate but differ in task, data representation, optimization target, and evaluation regime.

2. Clinical prediction setting and data basis

The longitudinal forecasting version of L2C-TabPFN is evaluated on the TADPOLE Challenge dataset derived from ADNI. The study uses only the D1 and D2 splits. D1 is the training set with longitudinal data from ADNI1, ADNI GO, and ADNI2, while D2 is the rollover/prediction set used for forecasting evaluation. D3 is cross-sectional single-visit data, and D4 contains rollover participants with at least one of the outcomes, but neither D3 nor D4 is used in the reported experiments (Ding et al., 25 Aug 2025).

The prediction targets are heterogeneous. DX is a multiclass classification task over cognitively normal (CN), mild cognitive impairment (MCI), and probable Alzheimer’s disease (AD). ADAS-Cog is a regression target corresponding to the Alzheimer’s Disease Assessment Scale–Cognitive Subscale. Ventricular volume is a regression target defined as ventricle volume normalized by intracranial volume, and is treated as a key proxy for neurodegeneration and brain atrophy (Ding et al., 25 Aug 2025).

The underlying variables are multimodal and include demographics, APOE4 genotype status, sex (is_male), education (educ), marital status, current age, months since baseline (month_since_baseline), diagnosis, clinical assessments such as MMSE and CDR-related measures, and imaging biomarkers including ventricles, fusiform, whole brain, hippocampus, middle temporal, and intracranial volume. The methodological difficulty is that these records are longitudinal, irregular, multimodal, and variable-length, so a patient’s raw history cannot be passed directly to a fixed-width tabular predictor (Ding et al., 25 Aug 2025).

3. Longitudinal-to-cross-sectional construction

The L2C transformation converts each patient’s history into a fixed-length feature vector at each prediction time point by summarizing all information available up to, but not including, the current time. This is presented as a way to handle irregular sampling, missing visits, and variable follow-up lengths while making each forecasting instance a cross-sectional snapshot that retains information about both current status and prior trajectory (Ding et al., 25 Aug 2025).

For each patient, observations are sorted by month since baseline. For a given time point $t$, starting from the second record, the method aggregates all previous observations $s<t$ for each numeric variable $x$. The summaries include the most recent measurement,

$mr_x = x(t^*), \qquad t^* = \max\{s \mid s < t\},$

the time since the most recent measurement,

$\Delta t_{mr} = t - t^*,$

and a change-rate statistic based on the two most recent prior observations,

$$mr\_change_x = \frac{x(t^*) - x(t^{**})}{t^* - t^{**}}, \qquad t^{**} = \max\{s \mid s < t^*\}.$$

It also records the minimum and maximum historical values and the times since those extrema:

$$low_x = \min\{x(s) \mid s < t\}, \qquad \Delta t_{low} = t - \min\{s \mid x(s) = low_x\},$$

$$high_x = \max\{x(s) \mid s < t\}, \qquad \Delta t_{high} = t - \min\{s \mid x(s) = high_x\}.$$

For diagnosis, the representation is augmented with progression-oriented indicators such as the most recent diagnosis, a “milder” state flag, and times since best and worst diagnosis states (Ding et al., 25 Aug 2025).

After summarization, each transformed instance contains longitudinal summary features, demographic variables, and the current outcome targets. The paper further applies data augmentation following the Frog line of work: if a patient has $n$ visits, then for each cutoff $t_m$ with $s<t$0, the model is trained to predict all later visits $s<t$1, yielding

$s<t$2

training examples. Because visit spacing is irregular, month_bl is replaced or augmented with a horizon feature encoding the interval from the cutoff time to the target time. During forecasting, the model sweeps over horizons by updating horizon and time_since_mr_XXX while leaving unaffected summary features unchanged (Ding et al., 25 Aug 2025).

4. TabPFN integration and experimental procedure

TabPFN is the downstream predictor after L2C conversion. It is described as a transformer-based tabular foundation model designed for small to medium-sized tabular datasets, roughly up to 10,000 samples and 500 features, and pretrained on millions of synthetic tabular datasets generated by a structural causal model. In this application, its appeal is that it can operate on the post-L2C fixed-length vectors, handle missing values, heterogeneous scales, and outliers, and perform in-context learning without additional gradient-based training on the downstream task. The paper describes TabPFN as using distinct embeddings for table cells and a two-way attention mechanism consisting of inter-feature attention within a sample and inter-sample attention across samples, with cached keys and values for training samples to enable efficient single-forward-pass inference (Ding et al., 25 Aug 2025).

The pipeline is: preprocess the ADNI/TADPOLE table, sort visits by time, apply L2C to generate fixed-length vectors, augment with horizon-based examples, and feed the resulting tabular vectors into TabPFN to predict DX, ADAS-Cog, or Ventricles. After preprocessing and augmentation, the reported total train sizes are 7,340 for DX, 7,075 for ADAS-Cog, and 4,990 for Ventricles. Validation folds are constructed by patient so that no patient appears in both training and validation, and the validation procedure uses the first half of each patient’s history to predict the second half. Evaluation uses MAUC and BCA for DX, and MAE for ADAS-Cog and Ventricles (Ding et al., 25 Aug 2025).

Hyperparameters for TabPFN are tuned with Optuna over 90 trials, using StratifiedKFold for classification and KFold for regression. The search space includes n_estimators from 1 to 31 in steps of 2, softmax_temperature from 0.1 to 2.0 on a log-uniform scale, and average_before_softmax as a Boolean. The selected settings are n_estimators=25, softmax_temperature=1.981, and average_before_softmax=True for DX; n_estimators=9, softmax_temperature=1.212, and average_before_softmax=True for ADAS-Cog; and n_estimators=31, softmax_temperature=0.718, and average_before_softmax=True for Ventricles. The principal baseline is Frog, an XGBoost-based method using the same L2C transformation and representing the state-of-the-art benchmark from prior work (Ding et al., 25 Aug 2025).

5. Empirical findings and interpretability

The central quantitative comparison is against Frog on the D2 evaluation set, averaged over five random seeds. The reported results are as follows (Ding et al., 25 Aug 2025).

Task	L2C-TabPFN	Frog
DX (MAUC)	0.9138 ± 0.0007	0.9258 ± 0.0003
DX (BCA)	0.7839 ± 0.0028	0.8056 ± 0.0019
ADAS-Cog (MAE)	6.7932 ± 0.0473	6.0030 ± 0.0229
Ventricles (MAE)	0.1577 ± 0.0005	0.1781 ± 0.0008

These numbers establish a task-dependent profile rather than a uniform superiority claim. Frog is better on both DX metrics and on ADAS-Cog MAE. L2C-TabPFN is better on Ventricles MAE, and the paper identifies ventricular volume prediction as the standout result, describing it as a statistically significant improvement and the method’s state-of-the-art outcome. The paper further notes that ADAS-Cog forecasting is historically difficult and may be weak as a benchmark, which qualifies the comparison on that target (Ding et al., 25 Aug 2025).

Interpretability is examined with SHAP. For the Ventricles model, the most important features are mr_Ventricles, high_Ventricles, low_Ventricles, together with high_WholeBrain, mr_WholeBrain, and ICV-related measures. For the DX model, the dominant features are mr_DX, high_CDRSB, best_DX, and worst_DX. For the ADAS model, importance is spread across a mixed set of anatomical measures, cognitive scores, demographics, and APOE4. The paper describes Frog’s SHAP profile as sparser and more concentrated on fewer top features, whereas TabPFN distributes importance more broadly across features; this is presented as a possible reason that TabPFN may capture subtler multimodal interactions, particularly in ventricular forecasting (Ding et al., 25 Aug 2025).

6. Limitations, adjacent literature, and broader significance

Several limitations are explicit. The method depends on handcrafted L2C feature engineering rather than on a native sequential architecture, and its empirical profile is target dependent: the strongest benefit appears in ventricular volume regression, while diagnosis and ADAS-Cog remain better served by Frog in the reported experiments. The study also uses only TADPOLE D1 and D2. The stated future direction is to extend evaluation to additional longitudinal datasets and to further explore predictive behavior and interpretability in real-world clinical settings (Ding et al., 25 Aug 2025).

L2C-TabPFN also sits within a broader biomedical and methodological TabPFN literature. A separate Alzheimer’s study on 3-year MCI-to-AD conversion, using biomarker-only multimodal features from TADPOLE/ADNI, reported TabPFN holdout AUC $s<t$3 overall and AUC $s<t$4 at $s<t$5, emphasizing advantages in low-data regimes rather than longitudinal multi-target forecasting (Ye et al., 29 Apr 2026). TabPFN-specific interpretability work has adapted ICE, PD, ALE, LOCO, Kernel SHAP with exact retraining, and data valuation to the context-based inference regime of TabPFN, making feature and training-sample analysis feasible without conventional retraining (Rundel et al., 2024). In low-label tabular learning, LUCoS proposes cold-start, label-free context selection for TabPFN / TabPFN-2.5 by embedding unlabeled rows with an unsupervised PFN encoder and selecting latent-space medoids, arguing that context composition strongly determines downstream performance (Ipas et al., 26 May 2026). Meanwhile, the causal-discovery paper that reuses the name L2C-TabPFN freezes TabPFNv2, extracts embeddings from the first $s<t$6 layers, and decodes causal graphs with $s<t$7 universal causal tokens aggregated to $s<t$8 representative tokens under a binary cross-entropy objective plus an acyclicity constraint (Swelam et al., 10 Nov 2025).

Taken together, these works place the longitudinal L2C-TabPFN in a family of systems that repurpose a frozen or lightly adapted tabular foundation model through task-specific context design, representation transformation, or lightweight adapters. A plausible implication is that the defining feature of the longitudinal L2C-TabPFN is not the acronym alone, but the specific combination of handcrafted temporal summarization, horizon-aware augmentation, and TabPFN-based forecasting for Alzheimer’s disease progression.