Predicting Depression Risk in Patients With Cancer Using Multimodal Data: Algorithm Development Study.

Area of Science:

• Oncology
• Psychiatry
• Data Science
• Machine Learning

Background:

• Cancer patients undergoing systemic treatment frequently experience depression.
• Early identification of at-risk individuals is crucial for timely intervention.
• Prediction models can aid healthcare professionals in identifying vulnerable cancer patients.

Purpose of the Study:

• To develop and evaluate a predictive model for depression risk in cancer patients within the first month of treatment.
• To explore the utility of machine learning and natural language processing for depression risk prediction.

Main Methods:

• Utilized electronic health record data and unstructured text (patient emails, clinician notes) from 16,159 cancer patients.
• Developed multimodal prediction models using least absolute shrinkage and selection operator (LASSO) logistic regression and Bidirectional Encoder Representations from Transformers (BERT).
• Assessed model performance using area under the receiver operating characteristic curve (AUROC), calibration curves, and decision curve analysis.

Main Results:

• The best-performing models, LASSO logistic regression using structured data and email classification scores, achieved an AUROC of 0.74.
• Bidirectional Encoder Representations from Transformers (BERT) models showed moderate performance (AUROC ~0.71).
• Models based solely on clinician notes or email classification scores performed poorly; risks were underestimated for female and Black patients.

Conclusions:

• Machine learning and multimodal models show promise for predicting depression risk in cancer patients.
• Limitations include potential biases and the need for further model refinement and validation.
• Future research should focus on improving outcome labels, predictors, and addressing subgroup biases.