Source-linked AI summary
Data Provenance and Ethics in the Curricula of U.S. Health AI Education Programs
Raffaele Giancotti, Francesca Santucci, Amama Anjum, Monique de Villa, Rupsa Mitra, Soham Biren Katlariwala, Vishala Mishra, Leo Anthony Celi
TL;DR
Clinical AI failures have often originated in how training data were generated, while whether health AI programs teach practitioners to examine provenance has been unclear. The authors analyzed advertised curricula at U.S. universities and found that ethics was widely covered but provenance and sustainability were not. Their findings describe the surveyed programs, with advertised curricula an imperfect proxy for classroom practice.
Problem
Whether health AI education programs teach practitioners to examine data provenance is unknown, while prior curricular reviews rarely treat provenance as a distinct competency.
Source: Introduction · pp. 1–2Method
The authors cross-sectionally coded publicly advertised curricula and instructor information for 36 U.S. university health AI programs identified through a structured multisource search.
Source: Introduction · p. 2, Data Extraction · p. 3Results
Ethics was covered widely, but provenance was rare and no program addressed sustainability; most named instructors came from computational disciplines.
Source: Discussion · p. 8, Conclusions · p. 9Takeaways & Limitations
The surveyed programs’ curricula widely taught ethics but rarely taught how clinical data are produced, and none taught computation’s environmental cost.
Source: Conclusions · p. 9Advertised curricula imperfectly represent classroom practice, and programs silent in public materials were counted as not covering a domain.
Source: Discussion · p. 8
Abstract
from arXiv · showhide
Failures of deployed clinical artificial intelligence (AI) systems have often originated in how their training data were generated. Whether the programs that train the health AI workforce teach practitioners to examine data provenance is unknown. We conducted a cross-sectional analysis of the publicly advertised curricula of health AI education programs at U.S. universities, identified through a structured multisource web search between April and July 2026. Two investigators coded coverage of data provenance (structural, dataset limitations, or data management only), ethics, and sustainability; collaborative structure; technical prerequisites; and the disciplinary background and gender of named instructors. The primary outcome was provenance coverage at the structural or dataset-limitations level. Of 43 candidate entries, 36 programs at 25 institutional units in 21 universities were included. Ethics was covered by 29 programs (80.6%; 95% confidence interval [CI], 65.0 to 90.2), data provenance by 8 (22.2%; 95% CI, 11.7 to 38.1), and structural provenance by 6 (16.7%; 95% CI, 7.9 to 31.9). None of the seven master's degree programs met the primary outcome, and no program addressed sustainability. Eighteen programs included team-based work, but only 5 assembled teams across clinical and technical backgrounds. Of 92 named instructors, 76.1% came from computational disciplines and 3.3% from the social sciences or humanities; among the 51 whose gender could be determined, 66.7% were men. U.S. health AI education programs widely taught ethics but rarely taught how clinical data are produced. None addressed the environmental cost of computation, and most instructors came from computational disciplines.
Introduction
Failures of clinical AI have been traced to how training data were generated and what labels measured, making data provenance a relevant competency for health AI practitioners. Yet curricular reviews rarely distinguish provenance from ethics, leaving its teaching in health AI programs unclear.
Introduction
- Clinical AI failures traced upstream to training-data origins and assumptions, including a poor proxy label that led to systematic underestimation of Black patients’ needs.Investigators uncovered these problems by asking where data came from and what they measured.Source: Introduction · pp. 1–2
- Prior reviews report limited AI ethics coverage in medical education and rarely treat data provenance as a distinct competency [5].Some reviews fold dataset representativeness into ethics, so ethics coverage alone cannot establish provenance teaching.Source: Introduction · p. 2
- This study examines 36 U.S. university programs for provenance, ethics, sustainability, collaboration, prerequisites, and instructor backgrounds.The programs were prominent but not randomly sampled; the authors present the findings as a baseline for annual repetition.Source: Introduction · p. 2
Materials and methods
The study cross-sectionally analyzed publicly advertised curricula of U.S. university health AI programs identified through a structured multisource search. Investigators coded program content and instructor characteristics from public documents, then summarized coverage descriptively.
Study Design
- The cross-sectional analysis examined publicly available program documents and did not undergo institutional review board review.The review was not registered and had no advance protocol.Source: Study Design · p. 2
Search Strategy
- Candidates came from Google searches, two retrieval-augmented language-model assistants, and hand searches of prior AI-education reviews.Every candidate was verified against an official institutional webpage, and coded content came from primary sources.Source: Search Strategy · pp. 2–3
Eligibility and Selection
- Of 43 candidate entries, the final sample comprised 36 programs at 25 institutional units in 21 universities after duplicates and ineligible offerings were removed.Eligibility required a U.S. university setting, substantial health AI focus, an open offering, and published curriculum and audience descriptions.Source: Eligibility and Selection · p. 3
Data Extraction
- Investigators extracted program format, audience, prerequisites, collaboration, domain coverage, and instructor characteristics, retaining unreported items in denominators.Instructors were counted once; discipline came from public biographies’ terminal degrees and appointments, and gender from biography pronouns.Source: Data Extraction · p. 3
Definitions
- Provenance meant conditions of data generation and their limits on model learning, coded as structural, dataset-limitations, or data-management-only coverage.The primary outcome combined structural provenance and dataset limitations; ethics coding accepted any public mention of ethics-related topics, while sustainability required mention of environmental, energy, or computational cost.Source: Definitions · pp. 3–4
Coding and Statistical Analysis
- Two investigators independently coded all programs and resolved disagreements by consensus; results were reported as counts and percentages with Wilson 95% confidence intervals.Chance-corrected agreement could not be computed because pre-consensus assignments were not retained; the sample was not a probability sample, and no hypothesis tests were performed.Source: Coding and Statistical Analysis · p. 4
Results
Across 36 programs, ethics coverage was common while provenance coverage was uncommon, and no program addressed sustainability. Named instructors were predominantly from computational fields, while cross-disciplinary team structures were rare.
Programs and Instructors
- 76.1% of 92 named instructors came from computational disciplines, compared with 3.3% from social sciences or humanities.Gender could be determined for 51 instructors, of whom 34 (66.7%) were men; the table concerns disciplinary background and gender.Source: Programs and Instructors · p. 6, Table 1 · p. 6
Collaboration and Admission
- 18 programs (50.0%) listed team-based work, but only 5 of those 18 (27.8%) explicitly assembled clinical–technical teams or paired learners with clinical mentors.Ten programs (27.8%) required prior technical coursework or proficiency for admission.Source: Collaboration and Admission · pp. 6–7, Table 2 · p. 7
Responsible AI Domains
- 29 programs (80.6%) covered ethics, while 8 (22.2%) met the primary provenance outcome and none addressed sustainability.Structural provenance appeared in 6 programs (16.7%); dataset limitations alone appeared in 2 (5.6%), and 3 additional programs taught data management only.Source: Responsible AI Domains · p. 7, Table 3 · p. 7
Discussion
Across 36 U.S. health AI programs, ethics was common while teaching about how clinical data come to exist was rare, and no program addressed computation’s environmental cost. The discussion links this pattern to curricula’s technical framing and notes limits in what advertised materials can establish.
Discussion
- Ethics appeared in four of five programs, but provenance teaching appeared in about one in five and structural provenance in one in six; no program addressed computation’s environmental cost.Provenance appeared mainly in single courses and in none of the master's programs.Source: Discussion · p. 8
- The authors interpret the pattern as a computer-science-derived template that treats health data as neutral inputs, placing ethics after the technical core and leaving computational costs outside the discipline.They argue education shapes which questions practitioners learn to ask, while also noting that safety and equity require audits, regulation, procurement standards, and community governance.Source: Discussion · p. 8
- Three quarters of named instructors came from computational disciplines, and only five programs used collaboration to connect model builders with people who had observed clinical data being made.More than a quarter of programs required technical preparation, which the authors say may screen out clinicians, community health workers, and social scientists.Source: Discussion · p. 8
- HST.953 pairs students with intensive care clinicians to examine how record data were generated, while PHC 3621 devotes a semester to whose data are used and how bias can produce discriminatory algorithms.The authors propose publishing complete syllabi with learning outcomes and revisiting content on a fixed cycle; they intend the survey as the first in an annual series.Source: Discussion · p. 8
- Advertised curricula are an imperfect proxy for classroom practice, and the reported gaps may be lower bounds because socially valued topics can be overstated in marketing and silent domains were coded as absent.Provenance coverage ranged from 16.7% under the narrowest definition to 30.6% under the broadest; double-coding was performed, but agreement could not be quantified.Source: Discussion · p. 8
- The search was not exhaustive, favored programs with substantial web presences, and did not capture programs launched after July 2026; instructor details and gender were also unavailable or uncertain for some individuals.Named instructors were unavailable for six programs, and gender could not be determined for 41 of the 92 instructors.Source: Discussion · p. 9
Conclusions
Prominent U.S. health AI programs widely taught ethics but rarely covered how clinical data are produced. None addressed computation’s environmental cost, and three quarters of named instructors came from computational disciplines.
Conclusions
- Prominent U.S. health AI programs widely taught ethics but rarely taught how clinical data are produced.Source: Conclusions · p. 9
- None of the programs taught the environmental cost of computation.Source: Conclusions · p. 9
- Three quarters of named instructors came from computational disciplines.Source: Conclusions · p. 9
Supporting information
The supporting information includes supplementary tables, a screening register, a PRISMA-ScR checklist, and openly available coded materials.
Supporting information
- The supplementary appendix contains the full 36-program dataset, search queries and prompts, sensitivity analyses of provenance estimates, and a de-identified dataset of 92 named instructors.Source: Supporting information · p. 9
- The screening register lists all 43 candidate entries and their inclusion decisions and exclusion reasons.Source: Supporting information · p. 9
- The completed PRISMA-ScR checklist is provided as a separate file, and the coded dataset and codebook are openly available on Zenodo.Source: Supporting information · p. 9