Independent research · sole author · 2025 – present

Risk-aware healthcare cost prediction

A care team can only reach a small share of its members. The useful question is not what someone will cost on average, but who belongs in that small share.

Evaluation code on GitHub

the high-cost spending captured at top-0.5% capacity, against a calibrated logistic baseline (p = 0.018): 2.0×
quantile crossings in 745,560 pairs, against 1.54% for baselines: 0
expected calibration error: 0.006
manuscript numbers machine-verified against run outputs: 98 / 98

$19.9M

In budget terms

of additional high-cost spending captured per 100K enrollees at top-0.5% capacity

Illustrative and approximate. 95% CI $3.7M to $40.1M.

The question

Healthcare spending is heavy-tailed: a small number of people account for a large part of the total. Most cost models are tuned for the average. This research asks what a model should predict when capacity is fixed and only the top of the distribution can be acted on.

The model is a multi-task FT-Transformer in PyTorch. From 72,818 person-years of US national survey data (MEPS, 2018–2023) it predicts expected cost, 7 tail quantiles and nested high-cost risk in one model, and its quantiles never cross.

The shape of the model

One record per person-year

US national survey data, MEPS 2018–2023

One multi-task model

FT-Transformer, PyTorch

Expected cost
7 tail quantiles
Nested high-cost risk

What did not reach significance

The gain is significant across the top 0.5–2% of capacity. At the 5% endpoint the difference is +1.89 points with p = 0.072, and it is reported as non-significant. The +1.89 is the mean of the paired bootstrap differences, so it is not exactly the gap between the two plotted values (1.75).

Headline result

Share of total spending captured by the people flagged

model calibrated logistic baseline not significant

Flagging the top 0.5% of people, the model captures 6.24% of total spending against 3.11% for the baseline (2.01×, p = 0.018).

Count-based top-k% selection, model against a calibrated logistic baseline. Paired cluster bootstrap, B = 2,000.

Built to be checked

60 seeded, reproducible runs (15 core, 45 external) with 0 failed steps, and a 35-run ablation.
Survey-weighted paired bootstrap inference with B = 2,000.
The crossing comparison is against a bank of LightGBM quantile regressors and logistic classifiers, the strongest classical method for each task. On raw accuracy the two are close; the difference is that this model's outputs never contradict each other.
Rebuilt inputs are byte-identical to the originals: 13 of 13 file hashes match.
A 98-point automated check binds every number in the manuscript to hashed evidence.
Reported to TRIPOD+AI and PROBAST+AI standards.

Beyond the first paper

Transportability: the frozen model evaluated on 6 external datasets (up to 1.8M records), reporting where it transfers and where it does not.
Nowcasting: remaining-cost prediction within the year, under partial information.
Segmentation and allocation: risk segments derived from predicted cost distributions, and budget-constrained audits of competing allocation policies.

Manuscripts

Sole author, as Zehui Zhang. Not yet published, so each is listed by subject.

01
Multi-output healthcare cost prediction under capacity constraints: tail utility and consistent outputs
Prepared for submission to JAMIA
02
Transportability of a frozen healthcare cost model across external populations and data sources
In preparation
03
Within-year nowcasting of remaining healthcare cost under partial information
In preparation
04
Distributional segmentation of high-cost populations from predicted cost quantiles
In preparation
05
Allocating scarce care capacity under budget constraints: allocation policy against predictive targeting
In preparation

Patent applications (China, CNIPA; filed 2026, pending)

Joint prediction of medical expenditure with mutually consistent outputs.
Population stratification from predicted expenditure risk distributions.
Constrained allocation of care resources from predicted risk distributions.

The manuscripts are not yet published, so this page stays at the level of the question and the headline results. I am glad to walk through the methods in conversation.