EUVIMEDEUVIMED entdecken
Crossref

Abstract FR412: Tabular Foundation Models With Survey-Weighted Conformal Prediction for Forecasting Cost Escalation in US Adults With Hypertension: A MEPS Panel Study

Anshul Saxena, Sonal Yadav, Yanjia Zhang, Lili Zhou, William Hudson, Agasya Mukkapati, Muni Rubens

Hypertension · 2026-10 · 83 (Suppl_1)

DOI: 10.1161/hyp.83.suppl_1.fr412 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Hypertension, 2026-10. Verfasst von Anshul Saxena und 6 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Background

Hypertension drives much of US healthcare spending, yet predictive models typically give point estimates without statistically valid individualized uncertainty. The Medical Expenditure Panel Survey (MEPS) lacks measured blood pressure, so escalation is operationalized via cost/utilization, not clinical control. We compared a tabular foundation model (TabPFN) against a gradient-boosting baseline (LightGBM) for predicting year-2 high-cost escalation among adults with diagnosed hypertension, adding survey-weighted conformal prediction sets.

Methods

We used MEPS Panel 27 (2022 feature year to 2023 outcome year); 2,431 adults with self-reported hypertension and positive longitudinal weight met study inclusion criteria. Year-1 demographic, comorbidity, utilization, and antihypertensive-regimen features predicted top-quartile year-2 expenditure. Models used design-aware 5-fold cross-validation grouped by primary sampling unit; LightGBM used survey weights directly, while TabPFN used weighted-bootstrap resampling. Weighted AUC and Brier score were estimated with 95% CIs via stratified cluster bootstrap (BRR replicates unavailable for this weight). Isotonic calibration was assessed on a held-out split. Split conformal prediction, weighted and Mondrian (race/ethnicity, income), quantified individual uncertainty at 90% target coverage.

Results

TabPFN outperformed LightGBM on discrimination (weighted AUC 0.815 [95% CI 0.796-0.836] vs 0.785 [0.761-0.809]) and calibration (weighted Brier 0.136 [0.128-0.146] vs 0.155 [0.144-0.168]), non-overlapping CIs. Isotonic recalibration improved LightGBM (Brier 0.154->0.148) but slightly worsened already-calibrated TabPFN (0.136->0.138). Marginal conformal prediction reached near-target coverage for both models (TabPFN 91.1%, LightGBM 89.5%) with compact set size (1.2-1.3 of 2 labels). Mondrian calibration preserved aggregate coverage but exposed unstable per-subgroup coverage given limited calibration samples per group (~140-150).

Conclusions

A tabular foundation model achieved superior discrimination and calibration versus a gradient-boosting baseline for predicting cost escalation in a nationally representative hypertension cohort, while supplying valid, survey-weighted individual uncertainty via conformal prediction. Equitable subgroup coverage remains limited by calibration sample size, motivating larger group-conditional calibration sets.

Zitieren

APA

Saxena, A., Yadav, S., Zhang, Y., Zhou, L., Hudson, W., Mukkapati, A., & Rubens, M. (2026). Abstract FR412: Tabular Foundation Models With Survey-Weighted Conformal Prediction for Forecasting Cost Escalation in US Adults With Hypertension: A MEPS Panel Study. Hypertension, 83(Suppl_1). https://doi.org/10.1161/hyp.83.suppl_1.fr412

Vancouver

Saxena A, Yadav S, Zhang Y, Zhou L, Hudson W, Mukkapati A, et al. Abstract FR412: Tabular Foundation Models With Survey-Weighted Conformal Prediction for Forecasting Cost Escalation in US Adults With Hypertension: A MEPS Panel Study. Hypertension. 2026;83(Suppl_1). doi:10.1161/hyp.83.suppl_1.fr412

BibTeX
@article{Saxena2026,
  author = {Saxena, Anshul and Yadav, Sonal and Zhang, Yanjia and Zhou, Lili and Hudson, William and Mukkapati, Agasya and Rubens, Muni},
  title = {Abstract FR412: Tabular Foundation Models With Survey-Weighted Conformal Prediction for Forecasting Cost Escalation in US Adults With Hypertension: A MEPS Panel Study},
  journal = {Hypertension},
  year = {2026},
  volume = {83},
  number = {Suppl_1},
  doi = {10.1161/hyp.83.suppl_1.fr412}
}
Kostenlos starten →