Can we identify individuals at risk to develop multiple myeloma? A machine learning-based predictive model.

Area of Science:

• Hematology
• Medical Informatics
• Oncology

Background:

• Multiple myeloma (MM) often progresses asymptomatically, leading to organ damage by diagnosis.
• Electronic health records (EHR) offer potential for early risk identification.
• Predictive modeling can identify at-risk individuals before clinical manifestation.

Purpose of the Study:

• To develop and validate predictive models for identifying individuals at risk of developing multiple myeloma within five years.
• To leverage extensive EHR data for early MM risk assessment.

Main Methods:

• Retrospective analysis of EHR data from 2002-2019, comparing future MM patients with matched healthy controls.
• Development of an XGBoost model using >200 parameters and a simplified logistic regression model with 20 key variables.
• Validation of predictive models using a large cohort of MM patients and controls.

Main Results:

• Future MM patients exhibited distinct pre-diagnostic patterns, including elevated ESR, lower hemoglobin, and increased immune deficiencies.
• The XGBoost model achieved an AUC of 0.836 for predicting 5-year MM risk.
• A simplified logistic regression model demonstrated an AUC of 0.72 for individual risk prediction.

Conclusions:

• Machine learning models effectively predict the 5-year risk of multiple myeloma in asymptomatic individuals.
• These predictive tools, utilizing EHR data, can aid in early identification and intervention for at-risk populations.
• The models show promise for practical application in clinical settings for proactive MM risk assessment.