Get API key

Presage Breathing Model Card


THE FOLLOWING INFORMATION, AND PRESAGE'S SDK AND APP, ARE OFFERED FOR GENERAL WELLNESS AND INFORMATIONAL PURPOSES ONLY. NEITHER HAVE BEEN CLEARED BY THE FDA AND NEITHER MAY BE USED FOR MEDICAL DIAGNOSIS OR TREATMENT.

1. Model Details

Basic info: Presage vitals by video analysis generates breathing rate metrics from a video of a subject.

Organization: Presage Technologies

Model date: 2026-11-06

Model version: d205a6f0

Model type: A proprietary computer vision and signal processing pipeline that estimates breathing rate and breathing waveform from video of a subject.

License: The algorithm is currently proprietary, and licenses are granted with predefined agreement.

Contact: Questions can be sent to: support@presagetech.com


2. Intended Use

Model Uses

This breathing model was intended for use in the analysis and non-diagnostic utility of breathing mechanics. It was intended to be used with a video from a stationary device that contains the subject's face, chest and shoulders in view. It requires approximately 45 seconds of uninterrupted data. It is only intended to measure breathing rate values in the range of 4-40 breaths per minute.

Out-of-Scope Uses

As noted above. The Presage breathing model is not intended for diagnostic purposes. No alarms are provided and it is not an apnea monitoring or detection model.


3. Validation

Reference Standard: ETCO2 capnography via Biopac MP160. For each SDK breathing-rate timestamp t, the CO2 reference breathing rate is computed from CO2 peaks detected in a 30-second lookback window ending at t:

BRCO2 = 60 · (N − 1) / (tlast peak − tfirst peak)
where N is the number of CO2 peaks in the window, tlast peak is the time of the last peak, and tfirst peak is the time of the first peak. This produces a continuous ground-truth breathing rate aligned to each SDK output timestamp, rather than a simple count of peaks per window.

Comparison Methodology: Breathing measurements from the camera-based system were compared against time-aligned reference measurements. Ground truth signals were checked for quality using labeled signal annotations; segments with poor signal quality were excluded from analysis.


4. Data Demographics

Category Distribution
Total 93 subjects, 184 videos
Camera (videos) Samsung S24 Rear 24mm (tripod): 92, Logitech C920: 92
Sex (subjects) Female: 50, Male: 43
Age Group (subjects) 18-25: 29, 26-35: 39, 36-45: 12, 46-55: 6, 56-65: 6, 65+: 1
Fitzpatrick (subjects) Type 1: 14, Type 2: 14, Type 3: 7, Type 4: 24, Type 5: 21, Type 6: 13
Lighting (videos) Ring Light: 184

Reference standard: ETCO2 via biopac peak detection for breathing rate ground truth.


5. Data Provenance

Reference Instrumentation: Biopac research-grade physiological sensors: ETCO2 capnography (for breathing rate ground truth).

Camera Devices Tested: Samsung S24 Rear 24mm (tripod), Logitech C920 (tripod).

Average Camera Distance: Logitech C920: 36", Samsung S24 Rear 24mm: 39"

Data Handling: All subject data is de-identified. Derived metrics and anonymized identifiers are retained. Data is securely stored with access restricted to trained researchers.


6. Factors

The breathing metric model requires face and pose detection to identify the subject's chest region for motion analysis.

  • User: User's face and chest must be sufficiently visible and positioned within the imaging window.
  • Environment: Model is developed on images with various lighting and motion conditions.

These factors can affect model performance:

  • Lighting: Temporal changes in lighting affect feature tracking ability.
  • Clothing: Dark solid shirts in low lighting can challenge feature tracking.
  • Motion: Breathing model cannot tolerate large non-physiological motion or talking.
  • Chewing gum can degrade performance.

Lighting Conditions Tested:

  • Ring Light: Direct, front-facing illumination providing consistent lighting on the subject's face.

Other factors:

  • Instrumentation: Several cameras have been tested for efficacy.

7. Metrics

(at 80% Return Rate, Confidence >= 79)

  1. MAE: 0.38 BrPM
  2. RMSE: 0.70 BrPM

8. Quantitative Analysis

Computed vs Ground Truth at Confidence Thresholds

Breathing rate in 30-second windows, computed vs. ground truth, at 50% and 80% return and for all readings, colored by camera

Bland-Altman Plot

Breathing rate Bland-Altman plot at an 80% return rate: mean difference 0.13 BrPM, limits of agreement 1.47 and -1.21 BrPM

Confidence Lookup Table

Confidence >= MAE (BrPM) RMSE (BrPM) Pearson r Return Rate (%) N (samples)
0 0.74 1.98 0.876 100.0 4110
10 0.74 1.98 0.876 100.0 4110
20 0.73 1.94 0.880 99.9 4107
30 0.70 1.84 0.890 99.5 4091
40 0.65 1.67 0.907 98.3 4041
50 0.52 1.15 0.953 95.5 3924
60 0.48 0.93 0.969 93.5 3842
65 0.45 0.84 0.974 91.9 3778
70 0.43 0.77 0.979 89.2 3666
75 0.41 0.74 0.980 85.6 3517
79 0.38 0.70 0.982 80.0 3293
80 0.37 0.69 0.983 78.2 3215
85 0.30 0.57 0.988 60.6 2491
90 0.20 0.40 0.992 11.5 471

Performance

(at 80% Return Rate, Confidence >= 79)

By Camera Type

Camera Type N (samples) Return Rate (%) MAE (BrPM) RMSE (BrPM) Pearson r
Samsung S24 Rear 24mm (tripod) 1603 78.7 0.39 0.72 0.982
Logitech C920 1690 81.5 0.37 0.68 0.983

By Fitzpatrick Skin Type

Fitzpatrick N (samples) Return Rate (%) MAE (BrPM) RMSE (BrPM) Pearson r
Type I 406 70.6 0.50 0.89 0.977
Type II 454 75.5 0.32 0.50 0.990
Type III 248 82.4 0.34 0.50 0.982
Type IV 839 84.2 0.29 0.50 0.992
Type V 849 85.0 0.33 0.59 0.989
Type VI 497 77.9 0.59 1.09 0.912

Note: All Fitzpatrick types were tested under Ring Light only.

By Sex

Sex N (samples) Return Rate (%) MAE (BrPM) RMSE (BrPM) Pearson r
Male 1480 76.7 0.44 0.84 0.976
Female 1813 83.1 0.33 0.56 0.988

By Age Group

Age Group N (samples) Return Rate (%) MAE (BrPM) RMSE (BrPM) Pearson r
18-25 936 72.6 0.44 0.80 0.977
26-35 1343 83.5 0.38 0.72 0.981
36-45 554 90.1 0.30 0.49 0.984
46-55 226 72.0 0.40 0.72 0.986
56-65 213 82.6 0.26 0.49 0.992
65+ 21 84.0 0.21 0.25 0.957

By Lighting Type

Lighting N (samples) Return Rate (%) MAE (BrPM) RMSE (BrPM) Pearson r
Ring Light 3293 80.1 0.38 0.70 0.982

Confidence vs Breathing Waveform Correlation

Confidence >= Upper Waveform Pearson r Lower Waveform Pearson r Return Rate (%) N (videos)
0 0.795 0.614 100.0 184
10 0.795 0.614 100.0 184
20 0.795 0.614 100.0 184
30 0.795 0.614 100.0 184
40 0.798 0.615 99.5 183
50 0.808 0.614 97.8 180
60 0.828 0.639 93.5 172
65 0.839 0.646 91.8 169
70 0.844 0.657 88.0 162
75 0.848 0.658 83.2 153
80 0.857 0.643 72.8 134
85 0.876 0.642 49.5 91
90 0.933 0.684 1.6 3

Waveform Example

Waveform example Subject_076_TV_24mm_001, 20 to 50 seconds: SDK upper body breathing against the Biopac upper strain sensor and ETCO2 capnography
Waveform example Subject_076_TV_24mm_001, 20 to 50 seconds: SDK lower body breathing against the Biopac lower strain sensor and ETCO2 capnography

9. Fairness & Equity

Bias Assessment Methodology: Performance is stratified by Fitzpatrick skin type (I-VI), sex, camera type, and age group. Per-group metrics and Confidence averages are reported in the Quantitative Analysis tables above.


10. Ethical Considerations

As a remote sensing device, the risks posed to the subjects in the trial are minimal, including the association of each subject with corresponding biometric data. Mitigation of these risks include de-identifying all subject data, including videos, prior to saving it. Additionally, all data is securely stored with access to a select number of trained researchers.

The model is not intended for human life-critical decisions, diagnostics or prognostication.

11. Limitations and Tradeoffs

  • Does not work when the camera or phone is handheld; both the camera and the subject must be stationary
  • Requires at least 15 seconds of uninterrupted view of face and chest area
  • Currently only measures a single subject at a time
  • Talking during measurement invalidates breathing rate
  • A confidence value of 0 indicates the algorithm has not yet accumulated 30 seconds of data; no breathing rate estimate is produced until the 30-second warm-up period completes
  • Handheld (front-facing 27mm) camera data is excluded from this analysis because the camera must be stationary and the subject must remain still for accurate measurement

Download the full model card HTML, 3.4 MB