Wearable Device-Based Screening for Anemia Risk (WEAR-ANEMIA)

This is a prospective, observational, diagnostic accuracy study conducted at a single tertiary care hospital in Beijing, China. The purpose is to develop and evaluate a deep learning model for non-invasive binary classification of anemia using multimodal physiological signals acquired from a commercially available consumer wrist-worn wearable device.

Study procedures. Participants will be enrolled in two sequential stages. In the first stage, 250 participants will be enrolled for algorithm development and tuning. In the second stage, an additional 150 participants will be enrolled for independent testing of the locked model and for supporting iterative optimization. All participants will undergo wearable signal acquisition and venous blood sampling for hemoglobin measurement, which serves as the reference standard.

Signal acquisition differs by care setting. Hospitalized participants will undergo 24-hour continuous wearable monitoring and paired signal acquisition before and after a six-minute walk test. Outpatient participants will undergo short-duration resting acquisition with the device worn alternately on the left and right wrist. Outpatient participants may undergo up to three additional assessments if they return for routine clinical visits; these visits are opportunistic and recorded as they occur. Hospitalized participants undergo a single acquisition without follow-up.

Reference standard. Anemia is defined according to Chinese sea-level criteria: hemoglobin below 120 g/L in adult men, below 110 g/L in non-pregnant women, and below 100 g/L in pregnant women.

Sample size. The primary endpoint is diagnostic sensitivity. Assuming an expected sensitivity of 70%, an allowable error of 6%, and a confidence level of 95%, at least 225 participants with anemia are required. Allowing for approximately 10% dropout or unusable data, the anemia group is set at 300 or more participants. At least 90 participants with normal hemoglobin will be enrolled to estimate specificity. To support deep learning model training and generalizability assessment, the total target enrollment is 400 participants.

Statistical analysis. Model performance will be reported as sensitivity and specificity. Cross-validation will be used to assess model stability, and an independent test set will be used to assess generalizability.

The wearable device is non-invasive. Study participation does not alter any aspect of the clinical management of participants.