Model Details

Basic info:

Presage vitals by video analysis generates real time measures of relative blood pressure, hereto referred as phasic, metrics from a video clip containing a subject’s face and chest. The phasic signal is the trend in blood pressure changes of the subject over time, and does not belong to an absolute scale to be interpreted outside of comparing different points in time for a single video. An API has also been developed to allow users easy access to compute these metrics for commercial and scientific applications.

Organization developing model: Presage Technologies

Model date: 20250818t185657

Model version: 1.5.0-rc.3

Model type:The arterial waveform is extracted from video frames by first detecting the user’s face with MediaPipe BlazeFace, then applying proprietary image processing, neural networks, and signal processing to infer the waveform over time. The resulting arterial waveform contains relative measures of systolic (peaks), diastolic (troughs) and mean absolute blood pressure.

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

Where to send questions: Questions can be sent to: support@presagetech.com

Intended Use

Model uses:

This phasic model was intended for use by qualified clinicians and researchers for the analysis and non-diagnostic utility of cardiovascular mechanics. It was intended to be used with a video from a stationary device (such as a handheld, mobile or laptop camera), that contains the subject’s unobstructed face in view, and be of at least 10 consecutive seconds in length with at least 25 fps in frame rate.

Out-of-scope uses:

Presage phasic model is not intended for diagnostic purposes. Do not self-diagnose or self-medicate on the basis of the measurements. No alarms are provided, and it is not an hypotensive or hypertensive monitoring or detection model. It is currently not intended for use in highly dynamic environments, or with a highly moving camera. We ensure all users have acknowledged and agreed to our license agreement and terms of service for usage prior to use.

Factors

The relative blood pressure model first requires Mediapipe’s face landmarking algorithm to identify the face and 468 key feature points on the face (mesh model). Thus, if these features are not identifiable by Mediapipe’s algorithm, the metrics will not be calculable. Reference the model card here for extensive description on mesh detection factors:MediaPipe BlazeFace and MediaPipe Attention Mask.

The following factors can affect model performance:

User: User’s face orientation must be within angular parameters for detection by Mediapipe, off axis view decreases model’s detection abilities. User’s face and chest must be sufficiently visible, illuminated and positioned within the imaging window. " User’s face and body must be clear, and not blurry due to excessive motion or camera system. Users are instructed to maintain a fixed position. User demographics, including skin tone and facial hair, may impact return rate. Darker skin tone will require a higher threshold on illumination intensity, which if not met may reduce accuracy. Facial hair or other facial obstructions will reduce the area of analysis, which may affect accuracy.
Environment: Model is trained on images with various lighting, noise and motion conditions and with diverse augmentations. However, its quality can degrade in extreme conditions.

These factors can affect model performance:

Lighting: Temporal changes in lighting and illumination can affect the ability to correctly measure relative blood pressure.
Motion: Phasic model cannot tolerate large degrees of non-physiological subject motion. Phasic model cannot handle large degrees of camera motion, instability or periodic shakiness. It is recommended that both the camera be mounted and the user is in a fixed position before acquiring measurement.

Other factors:

Instrumentation: Several cameras, including mobile (Android), USB (Econ) and webcam (Logi) based devices with varying sensitivities and image resolutions have been tested for efficacy.

Metrics

Window-based Pearson’s correlation coefficient (r) of relative blood pressure. To calculate this, for every 30s time window that the model ingests, the output phasic waveform is predicted (RBP) and its peaks are evaluated (SBP), troughs (DBP) and mean of the peaks and troughs (MAP). The correlation between the predicted measures of relative systolic (SBP), diastolic (DBP), mean absolute pressure (MAP) and waveform shape (RBP) after subtracting the starting value are compared to a ground truth measure with equivalent processing. The mean of the correlation for all windows over all videos in the dataset is presented. Each 60s video is evaluated over 30s windows with a 1s stride, thus were 2280 samples total.
Video-based Pearson’s correlation coefficient (r) of relative blood pressure. To calculate this, for every 60s time window that the model ingests, the output phasic waveform is predicted (RBP) and its peaks are evaluated (SBP), troughs (DBP) and mean of the peaks and troughs (MAP). The correlation between the predicted measures of relative systolic (SBP), diastolic (DBP), mean absolute pressure (MAP) and waveform shape (RBP) after subtracting the starting value are compared to a ground truth measure with equivalent processing. The mean of the correlation for all videos in the dataset is presented. This value helps reflect the effect of model drift over a longer period of time.

Evaluation Data

The evaluation data consists of 39 videos. A clip of 60s from each video was run through the Presage phasic model for evaluation (either in 30s or 60s window sizes). Videos were acquired on users covering a range of demographic variability, including age, gender and Fitzpatrick scale. Three tripod videos were recorded simultaneously from a Samsung Android S24 mobile phone, e-con See3CAM-CU27, and a Logitech C920 webcam. A single handheld video was recorded separately with the front-facing camera of the Android S24 mobile phone. Corresponding physiological signals were recorded from a Biopac research grade CNAP non-invasive blood pressure cuff (NIBP).

Quantitative Data

Per 30s window r for relative systolic blood pressure: 0.22 with 95% CI: [0.22, 0.22]
Per 30s window r for relative diastolic blood pressure: 0.30 with 95% CI: [0.30, 0.31]
Per 30s window r for relative mean absolute pressure: 0.26 with 95% CI: [0.26, 0.27]
Per 30s window r for relative blood pressure: 0.67 with 95% CI: [0.67, 0.67]
Per 60s window r for relative systolic blood pressure: 0.20 with 95% CI: [0.20, 0.20]
Per 60s window r for relative diastolic blood pressure: 0.28 with 95% CI: [0.28, 0.28]
Per 60s window r for relative mean absolute pressure: 0.24 with 95% CI: [0.24, 0.24]
Per 60s window r for relative blood pressure: 0.65 with 95% CI: [0.65, 0.65]

Distribution of error figures:

Skin Tone (Fitzpatrick)	% of Dataset (num samples)	r SBP mean (30s)	r DBP mean (30s)	r MAP mean (30s)	r RBP mean (30s)
1	0.18 (420)	0.26	0.39	0.32	0.68
2	0.11 (240)	0.28	0.51	0.38	0.75
3	0.18 (420)	0.30	0.40	0.36	0.72
4	0.08 (180)	0.08	0.15	0.11	0.46
5	0.18 (420)	0.20	0.14	0.19	0.66
6	0.21 (480)	0.17	0.25	0.21	0.64

Skin Tone (Fitzpatrick)	% of Dataset (num samples)	r SBP mean (60s)	r DBP mean (60s)	r MAP mean (60s)	r RBP mean (60s)
1	0.18 (14)	0.24	0.35	0.29	0.67
2	0.11 (8)	0.23	0.42	0.31	0.73
3	0.18 (14)	0.26	0.41	0.35	0.70
4	0.08 (6)	0.24	0.22	0.25	0.46
5	0.18 (14)	0.15	0.15	0.17	0.63
6	0.21 (16)	0.13	0.19	0.15	0.62

Camera Type	% of Dataset (num samples)	r SBP mean (30s)	r DBP mean (30s)	r MAP mean (30s)	r RBP mean (30s)
Android	0.47 (1080)	0.22	0.32	0.27	0.68
Econ	0.53 (1200)	0.22	0.29	0.26	0.66
Logi	0.00 (0)	nan	nan	nan	nan

Camera Type	% of Dataset (num samples)	r SBP mean (60s)	r DBP mean (60s)	r MAP mean (60s)	r RBP mean (60s)
Android	0.47 (1080)	0.19	0.28	0.24	0.65
Econ	0.53 (1200)	0.20	0.28	0.25	0.65
Logi	0.00 (0)	nan	nan	nan	nan

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 in a HIPPA compliant database with access to a select number of trained researchers.

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

Presage Phasic (Relative Blood Pressure) Metrics Model Card