Figure 1: a.AnemoCheck smartphone app. The AnemoCheck app works by capturing an image of fingernail beds and calculating how pale the nailbed is. The paleness in the nailbed is correlated to a Hgb level estimate that is displayed on screen.b. Current smartphone app users and usage.The smartphone app used in this study is available in the US for both iOS and Android download. Map shows the locations of a random sample of 1,000 smartphone app users across the United States. The table highlights the data breakdown powering the algorithm, and number of app accounts and uses as of November 30, 2022 at 1:55PM ET.c. Personalized algorithm training and testing.In this publication, we characterized the effect of training algorithms on the individual level via both a clinically validated study and “real world” data collection. Users captured images of nailbeds on the same days as venous blood draws with blood Hgb levels that were determined by the current clinical standard, a clinical hematology analyzer. The app algorithm was trained by inputting blood Hgb level data with images of nailbeds taken on the same day (training). After the algorithm was trained, the user tested the newly customized algorithm (testing). Results from the customized algorithm were compared with blood Hgb levels taken or reported within 24 hours of app use to establish accuracy of the customized algorithm. Therefore, proof-of-concept was established for improving the app algorithm using personalized calibration in both clinical and real-world examples of personalized medicine.
Figure 2: Performance improves with customized trained app.A total of 40 subjects were enrolled in this study. Of that cohort, 35 subjects completed phase 1 of the study and had at least one phase 2 visit. 32 subjects completed both phase 1 and phase 2 and one subject was excluded. In the first four visits (Phase 1), images of fingernail beds were captured along with a venous blood draw for blood Hgb level.a. First, these data were compared against blood Hgb levels to serve as a control for app results without app training (n=35). Then, these data were used to train a custom algorithm for each subject.b. In visits five through eight of the study, the personalized algorithm was tested and compared with the results of venous blood draws taken during those visits (n=35). The average error of the personalized algorithm testing data points (visits 5-8) was 0.74 g/dL with a root mean square error of 0.97. Bias was 0.16 g/dL, which represented an improvement over the untrained app (n=35).c. App performance improves with increasing number of training data points (n=31). Subject data across 8 data points was analyzed assuming different numbers of training points and testing points. Average error, root mean squared error, and bias of the testing points are shown for the number of training data points. We have also included a data point in the table using the two lowest and two highest blood Hgb level data points to train the algorithm. The results indicate that the average error, root mean squared error and bias are all smaller when the algorithm is trained with data across a broader Hgb level range.
Figure 3: Real world use of the app validates clinical findings.
Real world app users and their self-reported data were used to complete app training and testing.a. First, 104 data points from 16 users who self-reported at least 5 lab tests were used to serve as a control for app results without app training. Comparing these data to the entered lab-derived Hgb levels resulted in a mean absolute error of the personalized algorithm testing data points was 0.71 g/dL with a root mean square error of 1.27 g/dL. Then, the 64 data points corresponding to each user first 4 entered laboratory Hgb results were used to train a custom algorithm for each user.b. The newly customized algorithm was tested with the next 40 data points and compared with the results of self-reported blood Hgb levels reported within 24 h of using the app. The mean absolute error of the personalized algorithm testing data points was 0.62 g/dL with a root mean square error of 0.85 g/dL, which represented an improvement over the untrained app in the “real world”.c. App performance improves with increasing number of training data points. Real world subject data across 8 data points was analyzed assuming different numbers of training points and testing points. Average error, root mean squared error and bias of the testing points are shown for the number of training data points.d. In total, 474 users entered 673 lab tests with Hgb levels > 10 g/dL. In these users, the mean absolute error and RMSE compared to the entered lab results was 0.61 g/dL and 1.1 g/dL, respectively.