Preprocessing to Address Bias in Healthcare Data.

Area of Science:

• Health Informatics
• Clinical Decision-Making
• Health Equity

Background:

• Multimorbidity, the presence of multiple chronic conditions, increases with age and impacts clinical decision-making.
• Underdiagnosis of multimorbidity in underserved populations, particularly rural patients, introduces bias into healthcare algorithms.
• Artificial intelligence (AI) systems may generate inaccurate predictions due to unknown conditions in patient data.

Purpose of the Study:

• To investigate bias in healthcare data concerning multimorbidity classification, specifically favoring metropolitan patients.
• To develop and test a preprocessing method to mitigate bias in AI models used for healthcare risk prediction.
• To evaluate the impact of bias removal on the performance of multimorbidity classification models.

Main Methods:

• Collected patient data from an academic hospital, analyzing multimorbidity trends in relation to rurality.
• Quantified bias in a classification model trained on diagnosis information, identifying discrimination against rural patients.
• Developed an unbiased training dataset and a new model, then tested its performance against unaltered validation data.

Main Results:

• The study observed a decrease in multimorbidity with increasing rurality, indicating a bias against rural patients in diagnosis data.
• A new model trained on preprocessed, unbiased data showed comparable performance to the initial model trained on biased data when tested on unaltered data.
• Preprocessing techniques effectively removed bias without significantly compromising classification performance.

Conclusions:

• Healthcare data preprocessing can successfully mitigate bias related to patient geography and underdiagnosis.
• Fairer risk prediction algorithms are achievable through bias correction, improving healthcare equity for underserved populations.
• Addressing data bias is crucial for the reliable and equitable application of AI in healthcare.