When Body Fat Percentage (BF%) shows variability across multiple scans under identical conditions, it raises concerns about accuracy—akin to the distrust we'd have for a household scale that gave inconsistent readings.
In the realm of healthcare biometrics, accuracy matters, especially when benchmarking against gold-standard technologies like DEXA scanners. My role focused on enhancing the user's likelihood of successful scans through improved guides and UX/UI enhancements, as well as investigating alternative data approaches, including multiple capture iterations and applying Kalman Filtering (KF).
Engineering and ML enhancements can often span several months and might yield an improvement of accuracy of 0.5%. However, there are "quick win" enhancements, potentially achievable within weeks, that could also impact accuracy significantly.
Alternative methods to refine scan outcomes include features like animated guides and confidence scores. Increasing the number of scans per session could theoretically yield a more accurate prediction. For instance, if three scans are performed and one is an outlier, discarding it in favor of the two consistent results could improve the overall accuracy—a process we could refer to as a "calibration" phase.
However, this approach is not without its drawbacks.
Bad data in = Bad data out
Three poor scans can also produce an inaccurate result. Our model is susceptible to bias – if two scans are inadequate but one is good, the accurate capture may be discarded as an outlier. We would first need to develop a confidence scoring system to rate each scan based on factors like lighting, pose, clothing, background, etc. This would help mitigate the bias risk.
We also ask users to perform three consecutive scans. Since this utilized the v1.x capture method of 10-second pauses between front and side photos, repeating three times extends the process. If a scan fails the positioning checks, requiring another attempt, it adds even more time.
There are also considerations around the capture approach itself. Is it better to do back-to-back scans in one session? Or space them across a day so users can take breaks? Given the need to change clothes, one session makes more sense.
Varying the setup each time could improve alignment diversity, but also risks distraction. If the first scan is difficult, it may negatively impact the experience and subsequent results. However, completing it once means they know what to do, likely improving each round.
In summary, we must weigh factors of bias mitigation, time to complete, user experience, and alignment variability. Confidence scoring and streamlining the process appear to be priorities for improvement. Further testing would reveal optimal timing and setup iteration.