Developing Equitable AI Diagnostics: A Technical Approach to Bias Mitigation

In this talk, we will explore the critical need for fairness in AI-driven healthcare, with a focus on mitigating bias in machine learning models. As AI systems become more integrated into healthcare diagnostics, addressing the disparities in model performance across diverse ethnic groups is paramount. This session will present a technical deep dive into the challenges of bias in medical imaging datasets and the resulting impact on healthcare outcomes for underrepresented populations.

We will begin by defining the types of bias commonly found in machine learning models, with a case study in skin cancer detection. We will demonstrate how training on imbalanced datasets exacerbates disparities in diagnosing skin cancer across different racial groups. Attendees will gain insight into practical techniques for rectifying these biases, including data augmentation, fairness-aware algorithms, and advanced evaluation metrics designed to assess model equity.

In addition to discussing technical solutions, we will also address the limitations and ethical considerations surrounding bias mitigation in healthcare AI, highlighting the importance of interdisciplinary collaboration in creating equitable diagnostic tools. By the end of the session, participants will be equipped with the knowledge to implement fairness techniques in their own AI models, promoting better outcomes for all patient populations.

Transcript

1. Developing Equitable AI Diagnostics: A Technical Approach to Bias Mitigation ODSC 2024 Accel AI Institute - - Presented by Laura Montoya, Accel AI Institute
2. Laura is a scientist and engineer turned serial entrepreneur and startup advisor. She leads tech social impact and ethical AI development as the founder and Managing Partner of Accel Impact Organizations, including Accel AI Institute and Latinx in AI (LXAI). Laura has led sessions on social impact, tech diversity, and ethical AI development for Creative Mornings, Katapult Future Fest, Silicon Valley Future Forum, Tech Inclusion Conference, Thrival Summit, Global Hive Summit, and keynoted the “Future of Work” for the Data and Society Conference at UC Berkeley. Laura has given guest lectures and technical workshops at Google, Santa Clara University Law, Stanford University Computational Social Science, and GTC Deep Learning School. Recently she spoke at TEDx Santa Barbara and has been featured in WITtalks and CIIS podcasts, Xconomy, Verizon News, and Forbes Leadership. Laura is a startup mentor with the IBM Hyperprotect Accelerator Program and has completed certification by Venture Capital University, a joint initiative of UC Berkeley Law, Venture Forward, and the National Venture Capital Association. Research Scientist Entrepreneur | Angel Investor Public Speaker Startup Mentor / Advisor Emerging Fund Manager Laura N Montoya Accel AI Institute ODSC 2024 -
3. 01. Perpetuation of Bias in Healthcare Diagnostics Table of Contents 02. Types of Biases, their origins, and mitigation techniques 03. AI for Melanoma Detection Case Study 04. PRISMA-EQUITY Recommendations 05. Pop Quiz
4. Mitigating Bias in Machine Learning Textbook Accel AI Institute ODSC 2024 - 03 “Towards rectification of machine learning bias in healthcare diagnostics: A case study of detecting skin cancer across diverse ethnic groups” Chapter Contribution By: Laura N Montoya and Jennafer Shae Roberts, Accel AI Institute
5. 01 Gain a comprehensive understanding of the risks associated with biases in technological healthcare applications 02 Understand the significance of mitigating harms in ML applications used at scale in healthcare, Where they have a direct impact on people’s lives and well- being. 03 Explain how biases can impact decision-making in healthcare technology and its potential consequences. Learning Objectives 05 Understand where to implement bias mitigation strategies in the ML lifecycle including in pre- processing, in- processing, and post- processing phases. 04 Identify and describe the following biases and how they apply to ML applications for healthcare: cognitive, evaluation, sampling, underestimation, and statistical biases. 04 ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - -
6. Bias in ML for Healthcare - healthcare bias ML learns from healthcare data ML bias bias in the data affects users of ML applications for data bias in ML applications leads to continued bias in heathcare Perpetuation of Bias in ML for Healthcare applications of ML in healthcare that contain bias ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - - 05
7. Population Data Feedback Loop Accel AI Institute ODSC 2024 -
8. BIAS ORIGINS Statistical Bias 01 03 02 Evaluation Bias Cognitive Bias 04 Sampling Bias 05 Underestimation Bias
9. ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - - Limitation Description Correlated Bias(es) Constrained time budget Time and cost required to train machine learning models accurately can be high Evaluation Bias, Sampling Bias, Statistical Bias, Labeling Bias Limitations of Machine Learning Models and Correlated Bias(es) - 08
10. ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - - Limitation Description Correlated Bias(es) Requires large datasets Unable to learn from limited training examples Sampling Bias, Statistical Bias Limitations of Machine Learning Models and Correlated Bias(es) - 09
11. ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - - Limitation Description Correlated Bias(es) Vanishing Gradient More layers than needed can lead to degradation of accuracy (saturation) Statistical Bias, Underestimation Bias Limitations of Machine Learning Models and Correlated Bias(es) - 10
12. ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - - Limitation Description Correlated Bias(es) Not generalizable Knowledge from one task can only be transferred to similar tasks Underestimation Bias Limitations of Machine Learning Models and Correlated Bias(es) - 11
13. ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - - Limitation Description Correlated Bias(es) Lacks Understanding No common sense knowledge of the world or the data it is being trained on Cognitive Bias Limitations of Machine Learning Models and Correlated Bias(es) - 12
14. ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - - Limitation Description Correlated Bias(es) Lacks Creativity or imagination Not useful for tasks beyond classification or dimensionality reduction on their own Cognitive Bias Limitations of Machine Learning Models and Correlated Bias(es) - 13
15. Definition and Cause Synonyms Lifecycle Occurrence Mitigation Methods Human bias, stemming from real-world inequities and discrimination, propagated through data, results in “unfair” predictions Social Bias, Historical Bias, Societal Bias, Individual Bias, Pre- existing Bias, Negative legacy, Healthcare Bias Can occur in all parts of the ML development lifecycle. Inclusion of diversity at all levels; education around the importance of inclusion and diversity Accel AI Institute ODSC 2024 Cognitive Bias -
16. Definition and Cause Synonyms Lifecycle Occurrence Mitigation Methods Built-in error, indicates the amount of which all observed values are wrong Bias of an Estimator, Bias Contribution, Algorithmic Bias, Technical Bias Can occur in all parts of the ML development lifecycle Increased overall size of dataset for training. Include a loss function, aka MSE Accel AI Institute ODSC 2024 Statistical Bias -
17. Accel AI Institute ODSC 2024 Statistical Bias - x w f(w*x) input output x w f(w*x + b) input output statistical bias b Perceptron Equation with Statistical Bias Perceptron Equation without Statistical Bias
18. Definition and Cause Synonyms Lifecycle Occurrence Mitigation Methods Dataset is non- representative due to limited availability of data, changing of population characteristics and behaviors over time, or false conclusions Representation Bias, Temporal Bias, Longitudinal Data Fallacy, Emergent Bias, Population Bias, Group Bias, Aggregation Bias, Behavioral Bias, Content Production Bias, (Self) Selection Bias, Availability Bias Data selection Retrain algorithm on updated Dataset which is representative of current use population Accel AI Institute ODSC 2024 Sampling Bias -
19. Definition and Cause Synonyms Lifecycle Occurrence Mitigation Methods Evaluation on a dataset that uses inappropriate, vague, non-descriptive, and/or non-inclusive performance metrics, thus is non- representative of the population Observer Bias, Funding Bias, Label Bias Data selection, model training and evaluation, Preprocessing (Label Bias) Use of inclusive datasets for training and testing the algorithm; Relabel data to match the truth Accel AI Institute ODSC 2024 Evaluation Bias -
20. Definition and Cause Synonyms Lifecycle Occurrence Mitigation Methods When a model underfits the data and is not able to generalize to unseen data Underfitting, Narrow Bias Can occur in all parts of the ML development lifecycle. Data collection, model training and evaluation, deployment, and continuous improvement. Increasing sample size of the minority group. Adding constraints: Utilizing cost-sensitive learning: Adjusting optimal threshold for minority group in post processing Accel AI Institute ODSC 2024 Underestimation Bias -
21. Accel AI Institute ODSC 2024 - DATA UNDERSTANDING DATA PREPARATION EVALUATION MODELING DEPLOYMENT MEDICAL UNDERSTANDING DATA DATA COLLECTION HISTORICAL BIAS MEASUREMENT BIAS DATA SELECTION REPRESENTATION BIAS SELF-SELECTION BIAS MEASUREMENT BIAS TEMPORAL BIAS SAMPLING BIAS EVALUATION BIAS MODEL TRAINING AND EVALUATION ALGORITHMIC BIAS SOCIAL BIAS EVALUATION BIAS STATISTICAL BIAS MODEL DEPLOYMENT POPULATION BIAS AGGREGATION BIAS USER-INTERACTION BIAS PREPROCESSING BIAS MITIGATION RESAMPLE PREVIOUS DATA INTRODUCE NEW DATA POSTPROCESSING BIAS MITIGATION MODIFY DECISION THRESHOLDS ADJUST MODEL OUTPUTS RELABELING IN-PROCESSING BIAS MITIGATION ADVERSARIAL DIBIASING DATA REGULARIZATION FAIRNESS CONTRAINTS
22. Accel AI Institute ODSC 2024 -
23. Accel AI Institute ODSC 2024 -
24. Case Study 23 ODSC 2024 Accel AI Institute - - ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - - Daneshjou, R., Vodrahalli, K., Liang, W., Novoa, R. A., Jenkins, M., Rotemberg, V., Ko, J., Swetter, S. M., Bailey, E. E., Gevaert, O., Mukherjee, P., Phung, M., Yekrang, K., Fong, B., Sahasrabudhe, R., Zou, J., & Chiou, A. (2021). Disparities in Dermatology AI: Assessments Using Diverse Clinical Images. Disparities in Dermatology AI: Assessments Using Diverse Clinical Images.
25. ODSC 2024 Accel AI Institute - - ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - - Case Study 24 Daneshjou, R., Vodrahalli, K., Liang, W., Novoa, R. A., Jenkins, M., Rotemberg, V., Ko, J., Swetter, S. M., Bailey, E. E., Gevaert, O., Mukherjee, P., Phung, M., Yekrang, K., Fong, B., Sahasrabudhe, R., Zou, J., & Chiou, A. (2021). Disparities in Dermatology AI: Assessments Using Diverse Clinical Images.
26. ODSC 2024 Accel AI Institute - - ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - - Case Study 25 Daneshjou, R., Vodrahalli, K., Liang, W., Novoa, R. A., Jenkins, M., Rotemberg, V., Ko, J., Swetter, S. M., Bailey, E. E., Gevaert, O., Mukherjee, P., Phung, M., Yekrang, K., Fong, B., Sahasrabudhe, R., Zou, J., & Chiou, A. (2021). Disparities in Dermatology AI: Assessments Using Diverse Clinical Images.
27. Towards Rectification of Machine Learning Bias in Healthcare Diagnostics: A Case Study of Detecting Skin Cancer across Diverse Ethnic Groups Accel AI Institute ODSC 2024 - Authored By: Laura N Montoya, Jennafer Shae Roberts, and Belen Sanchez, Accel AI Institute
28. Accel AI Institute ODSC 2024 - PRISMA GUIDELINES
29. Accel AI Institute ODSC 2024 -
30. Accel AI Institute ODSC 2024 -
31. Skin Tone Values and Hues 30 Accel AI Institute - ODSC 2024 VALUES DARK LIGHT MEDIUM HUES COLD NEUTRAL WARM OLIVE COLD NEUTRAL WARM OLIVE COLD NEUTRAL WARM OLIVE
32. Accel AI Institute ODSC 2024 -
33. Accel AI Institute ODSC 2024 -
34. Accel AI Institute ODSC 2024 -
35. Accel AI Institute ODSC 2024 -
36. Accel AI Institute ODSC 2024 -
37. Accel AI Institute ODSC 2024 -
38. ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - - Type of Bias Example: Melanoma case study Mitigation Methods in Melanoma Case Study Evaluation Bias Algorithm trained and tested on light skin, does not represent the whole population Use of datasets with diverse skin tones for training and testing the algorithm. Retraining with the DDI dataset. Statistical Bias Higher false negatives for cancerous cells on dark skin tones. Low accuracy of positive melanoma detection on dark skin. Increased size of the dataset for training. Retraining with the DDI dataset.Reduce statistical bias alongside variance in data by including a loss function, aka MSE. Cognitive Bias Historical studies done on only white individuals for skin cancer, dermatology textbooks rarely show non-white skin Inclusion of diversity at all levels; education around the importance of inclusion and diversity. Education for Health professionals and individuals regarding the prevalence of melanoma and implications for dark-skinned patients. Sampling Bias Previous datasets only represented light skin, omitting anyone with dark skin Retrain algorithm on Diverse Dermatology Dataset (DDI) Underestimation Bias Lack of fine-tuning to improve performance of algorithms on new data Finetune algorithms to demonstrate improved performance Types of Bias in Melanoma Detection and Mitigation Methods 37
39. Match questions 1-5 with answers a-e. Pop Quiz 38 What is sampling bias? 1. What is evaluation bias? 2. What is statistical bias? 3. What is cognitive bias? 4. What is underestimation bias? 5. a. When a model is trained and evaluated using a dataset that is non-representative b. Results from the dataset being non-representative of the population of intended use c. A built-in or naturally occuring error which indicates the amount of which all observed values are wrong d. Human bias which can result in predictions and decisions that are “unfair” e. When a model underfits the data and is not able to generalize to unseen data
40. Which are the recommended actions for mitigating bias in ML? (Select all true answers) Pop Quiz 39 Bias risks must be evaluated and proactively addressed at all stages. 1. Must document assumptions and decisions regarding ML applications. 2. Processes to discover bias proactively must be established during development and implementation. 3. Establish processes for tracking changes in social norms and the escalation of potential harm resulting from ML biases. 4. Include end-users in the co-development and prototyping of ML applications for transparency. 5. All of the above 6.
41. ODSC 2024 Accel AI Institute Developing Equitable AI Diagnostics - - Thank You www.accel.ai