A TABULAR FOUNDATION MODEL FOR OUTREACH TARGETING: PREDICTING SENIOR EMPLOYMENT PROGRAM WILLINGNESS AMONG OLDER NONPARTICIPANTS USING WELFARE SURVEY DATA
Author(s)
Miryoung Kim, PhD1, Hae Sun Suh, MA, MS, PhD2.
1College of Pharmacy, Sunchon National University, Suncheon-si, Korea, Republic of, 2College of Pharmacy and Institute of Regulatory Innovation through Science, Kyung Hee University, Seoul, Korea, Republic of.
1College of Pharmacy, Sunchon National University, Suncheon-si, Korea, Republic of, 2College of Pharmacy and Institute of Regulatory Innovation through Science, Kyung Hee University, Seoul, Korea, Republic of.
OBJECTIVES: Outreach resources for senior employment programs are finite, making it operationally important to identify which nonparticipants are most likely willing to join. We compared machine learning algorithms, including a tabular foundation model (TabPFN), to predict this willingness using nationally representative welfare survey data.
METHODS: Data came from Korean older adults not currently enrolled in a government-sponsored senior employment program (2024 Korean Study of Welfare and Living Conditions of Older Adults). The outcome was willingness to participate. We compared six algorithms—logistic regression, LASSO, random forest, histogram gradient boosting, XGBoost, and TabPFN—to predict this outcome. Unlike conventional algorithms, TabPFN generates predictions via in-context learning without dataset-specific fine-tuning. The primary feature set included sociodemographic, economic, health, and medical cost burden variables. Performance was assessed over 50 repeated stratified holdout splits. We evaluated discrimination (AUROC, AUPRC), calibration (Brier score, calibration slope), clinical utility (decision curve analysis), and outreach efficiency by examining the observed willingness rate among individuals ranked in the top predicted percentiles.
RESULTS: Among 3,000 nonparticipants, 814 (27.1%) expressed willingness to join. Across all six algorithms, AUROC ranged from 0.654 to 0.691 in the primary model, though differences were modest. TabPFN achieved AUROC 0.686 (95% CI 0.682-0.691) and calibration slope 0.972. Random forest had the highest AUROC (0.691) but a calibration slope of 1.901, indicating overconfident probability estimates that undermine reliable individual-level ranking. Targeting the top 5% of TabPFN-predicted scores identified respondents with 60.2% observed willingness—2.2 times the population rate. Economic variables, particularly household economic satisfaction and business income, were stronger predictors than health-related variables.
CONCLUSIONS: TabPFN achieved competitive discrimination and reliable probability estimates, supporting its use for outreach prioritization. Welfare survey data—particularly economic indicators—can identify older nonparticipants most likely to respond, enabling more efficient targeting of outreach efforts.
METHODS: Data came from Korean older adults not currently enrolled in a government-sponsored senior employment program (2024 Korean Study of Welfare and Living Conditions of Older Adults). The outcome was willingness to participate. We compared six algorithms—logistic regression, LASSO, random forest, histogram gradient boosting, XGBoost, and TabPFN—to predict this outcome. Unlike conventional algorithms, TabPFN generates predictions via in-context learning without dataset-specific fine-tuning. The primary feature set included sociodemographic, economic, health, and medical cost burden variables. Performance was assessed over 50 repeated stratified holdout splits. We evaluated discrimination (AUROC, AUPRC), calibration (Brier score, calibration slope), clinical utility (decision curve analysis), and outreach efficiency by examining the observed willingness rate among individuals ranked in the top predicted percentiles.
RESULTS: Among 3,000 nonparticipants, 814 (27.1%) expressed willingness to join. Across all six algorithms, AUROC ranged from 0.654 to 0.691 in the primary model, though differences were modest. TabPFN achieved AUROC 0.686 (95% CI 0.682-0.691) and calibration slope 0.972. Random forest had the highest AUROC (0.691) but a calibration slope of 1.901, indicating overconfident probability estimates that undermine reliable individual-level ranking. Targeting the top 5% of TabPFN-predicted scores identified respondents with 60.2% observed willingness—2.2 times the population rate. Economic variables, particularly household economic satisfaction and business income, were stronger predictors than health-related variables.
CONCLUSIONS: TabPFN achieved competitive discrimination and reliable probability estimates, supporting its use for outreach prioritization. Welfare survey data—particularly economic indicators—can identify older nonparticipants most likely to respond, enabling more efficient targeting of outreach efforts.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR173
Topic
Epidemiology & Public Health, Health Service Delivery & Process of Care, Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
No Additional Disease & Conditions/Specialized Treatment Areas