Adding accelerometry raises task-conditioned recoverability in wearable photoplethysmography

Technical Summary: Adding Accelerometry Raises Task-Conditioned Recoverability in Wearable Photoplethysmography

Problem Statement
Wearable photoplethysmography (PPG) is widely accessible but susceptible to degradation from motion, contact pressure, ambient light, and perfusion changes. A critical challenge in deploying PPG for downstream inference is distinguishing between two potential causes of poor model performance: model limitation (the observer failed to extract usable information) versus measurement limitation (the sensor acquisition failed to preserve task-relevant information). Predictive accuracy alone cannot resolve this ambiguity because the observed score conflates the properties of the measurement with the properties of the fitted model. This study addresses the need for a diagnostic that separates demonstrated performance from the theoretical information available within a specific measurement configuration.

Methodology
The study utilizes the Brno University of Technology (BUT) Smartphone PPG Database v2.0.0, comprising 82,966 balanced ECG-QRS event/non-event samples derived from 3,840 recordings across 38 subjects. The target task is distinguishing PPG windows centered on ECG-QRS reference events from matched non-event windows, using subject-grouped five-fold cross-validation to ensure independence.

The analysis employs the OptiCeil framework to evaluate two distinct metrics:

1. Demonstrated Performance: The highest balanced accuracy achieved by a "model zoo" consisting of logistic regression and histogram gradient boosting observers.
2. Task-Conditioned Recoverability: A comparative estimate of the maximum performance theoretically recoverable from the data. This is calculated using a nearest-neighbor geometry approach (evaluating one-nearest-neighbor error across raw, principal-component, and score-space representations) and transforming the error via the Cover-Hart relation.

The methodology treats recoverability as a falsifiable empirical quantity. If a stronger held-out observer outperforms a candidate recoverability estimate beyond a frozen tolerance, the candidate is discarded. The study performs two primary comparisons:

• Quality Stratification: Comparing expert-rated "good" versus "poor" recordings across eight perturbation conditions (coughing, finger movement, laughing, light change, pressure, rest, talking, walking).
• Sensor Fusion: Comparing PPG-only data against PPG combined with accelerometry (ACC) on the identical subset of ACC-capable records to isolate the effect of adding a physical measurement channel.

Key Results

• PPG-Only Baseline: For the ECG-QRS recovery task, the strongest held-out observer achieved a balanced accuracy (BA) of 0.5018, with a comparative recoverability estimate of 0.5454. The small gap (0.0436) suggests that under this specific representation and subject-held-out design, the smartphone PPG signal preserves little cross-subject discrimination information above chance.
• Expert Quality Correlation: Expert-rated "good" recordings demonstrated higher recoverability (0.5706) than "poor" recordings (0.5488). This ordering was consistent across all eight perturbation conditions, with the largest separation observed during laughing and pressure conditions. This validates the recoverability metric against an external, independent quality assessment.
• Impact of Accelerometry: On the identical record set, adding accelerometry increased the recoverability estimate from 0.5520 to 0.6148 (a gain of 0.0628). Concurrently, the achieved balanced accuracy rose from 0.5007 to 0.5957 (a gain of 0.0950). The descriptive gap between recoverability and achieved performance contracted significantly (from 0.0513 to 0.0191), indicating that the added sensor channel increased the available information state for the task.

Significance and Claims
The paper claims that task-conditioned recoverability serves as a distinct diagnostic tool that separates measurement limitations from model limitations. The results demonstrate that:

1. Measurement changes alter information states: Adding a physical sensor (accelerometry) to the same recording session increases both the performance an observer can reach and the theoretical recoverability of the task.
2. Actionable Engineering Guidance: The pair of "achieved performance" and "recoverability" clarifies the next engineering intervention. A high recoverability estimate with low achieved performance suggests a need for better observers or representation learning. Conversely, low values for both quantities indicate a need for acquisition improvements (e.g., better contact, additional sensors, or withholding inference).
3. Task Specificity: The findings are explicitly task-conditioned. While PPG showed limited recoverability for ECG-QRS event detection in this specific setup, the author notes that PPG may remain informative for other targets (e.g., pulse-rate estimation) or representations.

The study concludes that changing the physical measurement system and changing the model are distinct interventions. A measurement-facing recoverability diagnostic can help determine which intervention is required for a specific inference task.