Proposed TreatmentFormer model predicts rapamycin treatment status in Alzheimer's data
The abstract proposes TreatmentFormer, a three-stage deep learning framework that predicts rapamycin treatment status from brain imaging and tabular data in 23 participants; it reports about 71% average accuracy over 10 test runs and is a proposal, not a validated clinical tool. [1]
Automatic summary · not editorially reviewed
The key findings
The authors propose TreatmentFormer, a three-stage multimodal deep learning framework for predicting rapamycin treatment status from brain imaging and tabular data. [1]
In 23 participants with baseline and post-treatment timepoints, the model reached an average prediction accuracy of 71.25% across 10 independent test runs. [1]
A post hoc SHAP-based analysis identified biomarkers associated with treatment response, particularly in blood-based and inflammatory modalities. [1]
The source at a glance
- Study type / format
- Computational modeling study reported in an abstract
- Population / context
- 23 participants with baseline and post-treatment timepoints in the described cohort
- Summary based on
- Study abstract
- Publication status
- Preprint · peer-review status not confirmed
- Original publication
- August 26, 2026
Proposed framework for rapamycin data in Alzheimer's
The abstract notes that Alzheimer's disease has no known cure and that rapamycin, an approved inhibitor of the mTOR pathway, has shown promising anti-aging and neuroprotective effects. It says treatment effects are hard to characterize because biological systems interact and patient data are limited. The authors therefore propose TreatmentFormer, a three-stage multimodal deep learning framework for predicting rapamycin treatment status from brain imaging and tabular data such as microbiome profiles, blood-based biomarkers, cerebral blood flow measurements and clinical variables including gender, age and body mass index.
Supporting source [1]
Reported accuracy in 23 participants
The framework first uses Random Forest-based feature selection to reduce noise in high-dimensional tabular data, then modality-specific encoders align imaging and tabular inputs in a shared latent space via self-supervised contrastive learning, and a transformer-based architecture integrates the representations for treatment classification. Evaluated on a cohort of 23 participants with baseline and post-treatment timepoints, TreatmentFormer reached an average prediction accuracy of 71.25% across 10 independent test runs. The authors describe the performance as stable and consistent despite small sample size and heterogeneous data. A post hoc SHAP-based feature analysis identified key biomarkers associated with treatment response, particularly in blood-based and inflammatory modalities.
Supporting source [1]
Abstract-only scope and small-sample limits
The abstract itself names small sample size and heterogeneous data as challenges. The 71.25% figure comes from 10 independent test runs in 23 participants, so it describes prediction of treatment status in this dataset, not a demonstrated clinical benefit of rapamycin. The SHAP findings are post hoc associations, not evidence that these biomarkers cause treatment response. The authors call the approach promising for small-sample biomedical studies and suggest possible future uses such as predictive biomarkers and patient stratification; these are stated possibilities, not validated applications. Only the abstract was reviewed; the full text was not reviewed.
Supporting source [1]
How to interpret this
Abstract only; the full text was not reviewed. The reported accuracy comes from 23 participants and 10 test runs, and the biomarker findings are post hoc associations. The framework is proposed, not a validated clinical tool.
Based on the study abstract; the full paper, tables and supplementary data were not reviewed.
Source & supporting passages
Wang C, Woods C, Nguyen T, Liu J, Lin A, Cheng J. · europepmc.org · August 26, 2026