Data Provenance in Medical Machine Learning | Alexander Gibson

Alexander Gibson is a PhD student at the Queensland University of Technology and the Australian Centre for Health Services Innovation studying the intersection of metascience and clinical machine learning. One of his focus areas is data provenance, the Who, What, Where, When, Why, and How of datasets, and how neglecting this can lead to bad outcomes in medical machine learning not only in research, but also for clinical practice and medical device approval. 

CONTACT RANDY:

Feedback: metasciencematters@gmail.com

EPISODE LINKS:

Alex's preprint on unreliable diabetes and stroke datasets:

https://www.medrxiv.org/content/10.64898/2026.02.24.26347028v2

OUTLINE:

0:00 - Introduction

3:14 - The beginning of Alex's interest in clinical predictive modeling

5:05 - Alex's interest in metascience

6:42 - Choosing a dissertation topic/metrics hacking in machine learning

9:49 - Preprint on data provenance in medical datasets

12:33 - The diabetes and stroke datasets Alex investigated

16:46 - Major irregularities in the data

23:29 - TRIPOD+AI guidelines for auditing machine learning studies

25:26 - How unreliable studies can impact clinical practice and medical device patents

26:42 - Citation networks

27:37 - AI-generated formulaic medical machine learning studies

31:50 - Strategies for high-quality data provenance

33:53 - Patents citing unreliable studies, and how to integrate data provenance into peer review

35:23 - The biggest problems for clinical predictive modeling studies

37:02 - Resources and tools for improving rigor in machine learning

38:45 - Metrics reporting

40:45 - Choosing decision thresholds in predictive models

42:59 - The importance of clinical context in metrics reporting

45:21 - The unreasonable effectiveness of age and sex as predictors

47:53 - The roles of academia and industry in improving clinical machine learning studies

50:07 - Explanation versus prediction

52:51 - Advice and resources for students

54:27 - Outro