Amplitude Is Not Transferable: Perfusion-Driven Distribution Shift in Wearable PPG Signal Quality Assessment

Wearable photoplethysmography (PPG) is increasingly deployed for continuous monitoring in low-resource clinical settings. Automated signal quality assessment (SQA) is essential there: noisy windows must be discarded before any downstream physiological estimate can be trusted. But SQA models are usually built and validated on healthy volunteers. We report preliminary evidence that this practice fails systematically, and for a physiological reason, on the patients these systems are meant to serve. Signal quality assessment for PPG has converged on a fairly stable set of approaches. Template-matching methods compare each pulse against a subject-specific or population-average beat. Spectral methods score the concentration of power in the cardiac band. Statistical methods use skewness, kurtosis, and entropy as proxies for waveform regularity. These families are typically combined and passed to a classifier, and are typically developed on healthy volunteers or ambulatory recordings. Where clinical populations have been examined, the reported difficulty is usually motion artefact or probe displacement. These are problems of contact rather than of physiology. We built a lightweight SQA pipeline for a multi-wavelength wearable PPG device deployed for dengue monitoring at a tertiary infectious diseases hospital. We computed nineteen signal quality indices per sliding window, covering statistical, spectral, and pulse-morphology families. Mutual information selected the twelve most informative features. These fed a compact multilayer perceptron (1,029 parameters; 4.9 KB after int8 quantisation) sized for on-device inference on an ESP32-class microcontroller. Trained on healthy-volunteer recordings, the classifier marked 61% of held-out healthy windows as acceptable quality. On recordings from hospitalised dengue patients, that figure fell to 21-24%, and 66-71% of windows were marked probe-detached. The probe was worn continuously throughout, and this was documented. The cause is physiology, not sensor failure. In the critical phase of dengue, catecholamine-driven peripheral vasoconstriction collapses peripheral pulse amplitude. Median peak-to-peak amplitude fell from roughly 15,000 ADC counts in healthy recordings to a few tens of ADC counts in a patient recorded on illness day five. The most informative feature by mutual information was 95th-percentile absolute amplitude, the one feature that cannot transfer across perfusion states. The features ranked below it were relative band powers and waveform-shape descriptors, which are amplitude-invariant by construction. What we report is therefore different in kind from the failures usually described. The probe is attached and the pulse is present, but the amplitude scale on which the selected features were calibrated has shifted by nearly three orders of magnitude. Our contribution is not a better classifier but a diagnosis: feature selection optimised for informativeness within a single physiological state will, by construction, favour exactly the features least likely to survive a change in that state. This turns an apparent engineering defect into a distribution shift with an identifiable physiological driver, and it points to a specific fix: normalise amplitude within each recording, and rely on relative and morphological features. We are now testing this under leave-one-session-out validation on a larger patient cohort. We are also testing whether the same failure and the same fix carry over to a mechanically transduced pulse sensor, since optical and mechanical modalities both depend on peripheral pulsatility. These results are preliminary: three patient sessions, and window-level splits that make the reported accuracies optimistic. Session-level validation is underway. We report them because the direction and size of the effect look robust, and because the operational implication matters: quality control may be discarding data from the sickest patients.