The four-feature source-study comparison AI
Carlson et al. (2021) introduced the dataset and evaluated separately fitted BCG regressions for each participant and sensor. Their multiple-predictor models combined several features from one sensor. The maintained reproduction here focuses on Film0.
Four-feature regression
Natural-log I–J and J–K durations in seconds and J-minus-I/J-minus-K voltage amplitudes enter an ordinary least-squares model with an intercept. SBP and DBP are fitted separately for each person. Transformed features and pressure targets use trailing five-beat means. ECG detection is approximate and automatic; exact author-selected excerpts and manual exclusions are unavailable.
My same-recording fitted correlations average 0.518 for SBP and 0.502 for DBP; the paper’s Film0 values are 0.54 and 0.44. A fitted correlation uses data that also fitted the model. It is not a score for unseen later beats.
Matched evaluation in this project
The main report instead fits on early calibration windows and evaluates later endpoints 75–100. The four-feature model, 42-feature personal ridge and cohort-sharing model use the same valid pressure targets and evaluation windows. Their reported MAEs are this project’s held-out results, not publication accuracy claims.
The extensions change both features and estimator, so the comparison does not isolate the effect of feature count. This is an approximate reproduction and retrospective extension in a small, explored cohort. See feature definitions and validation context.
Carlson, C., Turpin, V.-R., Suliman, A., Ade, C., Warren, S., & Thompson, D. E. (2021). Bed-Based Ballistocardiography: Dataset and Ability to Track Cardiovascular Parameters. Sensors, 21(1), 156. doi:10.3390/s21010156. Dataset: doi:10.21227/77hc-py84.