Machine learning to identify precachexia and cachexia: a multicenter, retrospective cohort study.

Area of Science:

• Oncology
• Geriatrics
• Machine Learning in Medicine

Background:

• Early detection of precachexia is crucial for managing cachexia, but identification remains challenging.
• Precachexia detection is vital for effective prevention and treatment strategies.
• Current methods for identifying precachexia lack simplicity and efficiency.

Purpose of the Study:

• To develop a simple method for detecting cancer precachexia.
• To differentiate characteristics of precachexia from cachexia.
• To create accurate predictive models for precachexia and cachexia.

Main Methods:

• Utilized machine learning (ML) models trained on baseline characteristics of 3896 participants.
• Employed variable importance analysis to refine ML models for optimal performance.
• Validated model accuracy using receiver operating characteristic (ROC) values and calibration curves.

Main Results:

• Developed two logistic regression models for cachexia and precachexia detection with AUC values of 0.830 and 0.701, respectively.
• Identified key indicators for cachexia (eating changes, arm circumference, HDL, CAR) and precachexia (eating changes, serum creatinine, HDL, handgrip strength, CAR).
• Models demonstrated good accuracy and calibration, facilitating clinical application.

Conclusions:

• Successfully developed and validated ML models for identifying precachexia and cachexia.
• These models offer a practical tool for clinicians to improve early detection and diagnosis of precachexia.
• The findings support the integration of ML in clinical practice for nutritional assessment.