The proof-of-concept HeMonitor study showed that noninvasive hemoglobin measurement is “feasible,” according to machine-learning data presented at the Eleventh Annual Meeting of the Society of Hematologic Oncology.
Anne Kubasch, MD, of Leipzig University Hospital, and colleagues conducted the noninterventional, prospective-exploratory HeMonitor study to develop a noninvasive hemoglobin level prediction model based on photographs of the ocular conjunctiva and fingernails using machine learning.
They recruited two cohorts of patients, one that included patients with hematological malignancies who were undergoing regular hemoglobin measurements (n=373) and a second cohort that included volunteer healthy blood donors (n=188).
Dr. Kubasch and colleagues took photographs of the ocular conjunctiva and fingernails under laboratory conditions and gathered patient clinical data along with an invasively determined hemoglobin level that they measured the same day. They used this information to develop the training and testing dataset. The researchers then trained a Bayesian Ridge Regression model using the SciKit-Learn Python framework with the laboratory-derived hemoglobin measure as the target feature.
With the implemented approach, the researchers reported a mean hemoglobin deviation of 1.32 mmol/L when using photographs of conjunctiva compared to invasive hemoglobin measurement methods. They reported a mean hemoglobin deviation of 1.62 mmol/L when using photographs of fingernails compared to invasive hemoglobin measurement methods.
When they used photographs from conjunctiva in combination with photos of fingernails in a sequential prediction pipeline, the mean hemoglobin deviation was 1.42 mmol/L. However, there was “lower accuracy in patients below the anemia threshold (7.4 mmol/L) with frequent deviations of 4 mmol/L,” according to the study’s authors.
The proof-of-concept study showed that noninvasive hemoglobin measurement is “feasible,” but there were several limitations, according to the researchers.
“The main limitations of our results are lower accuracy in severely anemic patients, and thus, currently limited clinical applicability,” Dr. Kubasch and colleagues wrote. “Validation of our [machine-learning] model is currently ongoing within a larger, prospective study. Overall, by providing a way to monitor [hemoglobin] levels regularly at home, these measurements can contribute to better management of anemia and more personalized cancer care.”
Hefner S, Oeser A, Klötzer C, et al. HeMonitor: Machine Learning-Based Noninvasive Estimation of Hemoglobin (Hb) Value in Patients With Hematological Malignancies. Abstract MDS 337. Presented at the Eleventh Annual Meeting of the Society of Hematologic Oncology; September 6-9, 2023; Houston, Texas.