Post-processing methods for mitigating algorithmic bias in healthcare classification models: An extended umbrella review
Bibliographic record
Abstract
AI and predictive analytics have increased the speed of innovation in medicine. If left unchecked, however, algorithmic bias can exacerbate health disparities across race, class, or gender. Early bias mitigation literature has focused on addressing bias in the preparation and development phases of the algorithm life cycle (pre- and in-processing). Post-processing methods, applied at the point of implementation, are less computationally intensive and do not require re-building or training the model, allowing lower-resourced health systems to improve bias in off-the-shelf binary classification models, which are increasingly common within electronic medical records. This umbrella review sought to identify post-processing bias mitigation methods and tools applicable to binary healthcare classification models in healthcare and summarize bias reduction effectiveness and accuracy loss. This review was registered with PROSPERO and reported according to PRISMA 2020. PubMed and Scopus were searched in December 2023 for English-language reviews published post-2013 using an expanded search string from previous work on machine learning bias. Eligibility criteria followed the PICOT framework. Reviews were screened independently by two authors. Data were extracted from reviews using the Joanna Briggs Institute Extraction Form for Review of Reviews, as well as from cited studies (hence, an “extended” umbrella review). Quality was assessed using the Critical Appraisal Checklist for Systematic Reviews. Evidence was synthesized by mitigation method and effectiveness. Searches yielded 184 records. After duplicate removal, title/abstract, and full text screening, 11 reviews were included, citing 16 eligible studies. Post-processing methods tested included threshold adjustment (9 studies, cited by 8 reviews), reject option classification (6 studies, cited by 4 reviews), and calibration (5 studies, cited by 4 reviews). Threshold adjustment reduced bias across 8/9 trials; reject option classification and calibration reduced bias in approximately half of trials (5/8 and 4/8). Results were reported with heterogeneous fairness and accuracy metrics, making comparison difficult. A lack of effectiveness evaluation was noted across reviews. Four reviews identified 16 software libraries for addressing bias. Quality of the majority of reviews was weak due to inadequate reporting on methods. Threshold adjustment showed significant promise in post-processing bias mitigation for healthcare algorithms, followed by reject option classification and calibration. Future research should empirically compare post-processing methods on binary classification models using real-world healthcare data. As commercial algorithms proliferate, health systems require proven, achievable strategies to maximize fairness.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".