Probabilistic approach to longitudinal response prediction: application to radiomics from brain cancer imaging

TL;DR

This work tackles longitudinal response prediction from baseline radiomics in glioblastoma by introducing a probabilistic framework that models the intermediate outcome $Y_1^j$ via a Gaussian KDE across data splits and uses samples $\\widehat{\text{proba}}_1^j$ to drive a second-timepoint predictor $f_2^{L,i}$. The approach leverages only baseline features, providing explicit uncertainty quantification and avoiding dependence on intermediate follow-up radiomics, while remaining applicable to any number of timepoints. Across synthetic and Lumiere datasets, it achieves competitive performance with improved calibration and robustness, highlighting its suitability for sparse longitudinal imaging and potential generalization to other longitudinal imaging tasks. The methodology offers a principled path to incorporating uncertainty into longitudinal radiomics, with practical implications for timely decision-making in neuro-oncology and beyond.

Abstract

Longitudinal imaging analysis tracks disease progression and treatment response over time, providing dynamic insights into treatment efficacy and disease evolution. Radiomic features extracted from medical imaging can support the study of disease progression and facilitate longitudinal prediction of clinical outcomes. This study presents a probabilistic model for longitudinal response prediction, integrating baseline features with intermediate follow-ups. The probabilistic nature of the model naturally allows to handle the instrinsic uncertainty of the longitudinal prediction of disease progression. We evaluate the proposed model against state-of-the-art disease progression models in both a synthetic scenario and using a brain cancer dataset. Results demonstrate that the approach is competitive against existing methods while uniquely accounting for uncertainty and controlling the growth of problem dimensionality, eliminating the need for data from intermediate follow-ups.

Figures (5)

• Figure 1: All the patients included in the Lumiere dataset, with the response assessment at each timepoint. Pre-Op and Post-Op indicate pre- and post-operative timepoint; SD stable disease; PD progressive disease; CR complete response; PR partial response; nan indicates missing information about the timepoint; Post-Op/PD indicates second post-operative progression.
• Figure 1: Calibration curves for the baseline model to predict the response at first follow-up. The curves are generated by grouping predicted probabilities into bins and then plotting the average predicted probability for each bin against the corresponding observed frequency (fraction of positive cases). Top row: calibration curve of two model trained on Lumiere dataset. Bottom row: calibration curves of two models trained on the synthetic dataset.
• Figure 2: Subset of Lumiere dataset meeting the inclusion criteria, with response assessment at the first and second follow-up after baseline (pre-op label in the legend). For the benchmark experiment based on radiomics at the first follow-up, exclude patients $2$, $7$, $23$, $30$, $39$, $60$, $62$, and $89$.
• Figure 2: Examples of probability distribution of disease progression (response of class $1$) at the first follow-up. Light blue: normalized histogram (area $1$) of $\{\texttt{probas}_1^{i,j}\}_i$ obtained across splits, $i$, for each patient, $j$. Red: probability density function (PDF) of PD distribution estimated via G-KDE, $\hat{Y}^j_1$, for the $j$-th patient. Pink: samples drawn from the G-KDE PDF, $\{\widehat{\texttt{proba}}_1^{j}\}$. First column: examples from the synthetic dataset. Second column: examples from Lumiere dataset.
• Figure 3: Synthetic longitudinal dataset overview. Left panel: PCA representation of the synthetic dataset colored by the target label at first timepoint. Right panel: PCA representation of the synthetic dataset colored by the target label at second timepoint.