Development and early feasibility testing of machine-learning algorithms to non-invasively assess hemoglobin levels

Area of Science:

• Biomedical Engineering
• Hematology
• Artificial Intelligence in Medicine

Background:

• Anemia is a common complication in hematologic malignancies, impacting quality of life.
• Current hemoglobin (Hb) monitoring requires invasive blood draws, posing challenges for frequent assessment.

Purpose of the Study:

• To assess the feasibility and accuracy of non-invasive hemoglobin (Hb) estimation using image-based techniques and machine learning.
• To explore a combined framework integrating Hb estimates with patient-reported outcomes for home monitoring.

Main Methods:

• The HeMonitor study enrolled 367 patients with hematologic malignancies and 184 healthy donors.
• Fingernail and eyelid photographs were analyzed using Light Gradient-Boosting Machine (LightGBM) regression models for Hb prediction.
• A two-stage framework combined an image-based Hb predictor with a rule-based layer integrating EORTC Global Health and Fatigue categories.

Main Results:

• The best LightGBM model achieved a residual standard deviation of ±1.02 mmol/L for Hb prediction.
• Visual analyses indicated lower Hb levels correlated with impaired quality of life.
• The integrated framework showed concordance with clinician assessments, especially in borderline Hb ranges.

Conclusions:

• Non-invasive, image-based Hb assessment using machine learning is feasible in patients with hematologic malignancies.
• Combining digital biomarkers with patient-reported outcomes shows promise for patient-centered home monitoring.
• Prospective validation is necessary to confirm these findings for clinical application.