Predicting Patient Demographics From Chest Radiographs With Deep Learning

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Health equity and bias in AI

Background:

• Deep learning models are increasingly used in medical decision-making.
• Current AI models often lack diversity in training data, potentially leading to disparities in performance for underrepresented groups.
• Ensuring AI model fairness is crucial for equitable healthcare delivery.

Purpose of the Study:

• To assess the capability of deep learning models to classify patient demographics from chest radiographs.
• To evaluate model performance in predicting self-reported gender, age, ethnicity, and insurance status.
• To identify potential biases in AI models used in medical imaging.

Main Methods:

• Trained and tested deep learning models on 55,174 chest radiographs from the MIMIC-CXR database.
• Validated model performance using external datasets from CheXpert and a multihospital urban healthcare system.
• Evaluated model performance using macro-averaged area under the curve (AUC) metrics.

Main Results:

• Achieved near-perfect accuracy in predicting gender (AUC 0.999).
• Demonstrated high accuracy in predicting age (AUC 0.854-0.911) and ethnicity (AUC 0.854-0.911).
• Showed moderate accuracy in predicting insurance status (AUC 0.605-0.705).

Conclusions:

• Deep learning models can accurately predict patient demographics from chest radiographs.
• These models can be utilized to audit and enhance the diversity of training datasets.
• This approach can contribute to developing more equitable and robust artificial intelligence tools for diverse patient populations.