(231 days)
Tempus ECG-AF is intended for use to analyze recordings of 12-lead ECG devices and detect signs associated with a patient experiencing atrial fibrillation and/or atrial flutter (collectively referred to as AF) within the next 12 months. It is for use on resting 12-lead ECG recordings collected at a healthcare facility from patients:
- · 65 years of age or older,
- · without pre-existing or concurrent documentation of atrial fibrillation and/or atrial flutter,
- · who do not have a pacemaker or implantable cardioverter defibrillator, and
- · who did not have cardiac surgery within the preceding 30 days.
Performance of repeated testing of the same patient over time has not been evaluated and results SHOULD NOT be used for patient monitoring.
Tempus ECG-AF only analyzes ECG data. Results should be interpreted in conjunction with other diagnostic information, including the patient's original ECG recordings and other tests, as well as the patient's symptoms and clinical history. Tempus ECG-AF is not for use in patients with a history of AF, unless the AF occurred after a cardiac surgery procedure and resolved within 30 days of the procedure. It is not for use to assess risk of occurrence of AF related to cardiac surgery.
Results do not describe a person's overall risk of experiencing AF or serve as the sole basis for diagnosis of AF, and should not be used as the basis for treatment of AF.
Results are not intended to rule out AF follow-up.
Tempus ECG-AF is a cardiovascular machine learning-based notification software intended to analyze recordings of 12-lead ECG devices from patients 65 years of age and older. The software employs machine learning techniques to analyze ECG recordings and detect signs associated with a patient experiencing atrial flutter (collectively referred to as AF) within the next 12 months. The device is designed to extract otherwise unavailable information from ECGs conducted under the standard of care, to help health care providers better identify patients who may be at risk for undiagnosed AF in order to evaluate them for referral of further diagnostic follow up and address the unmet need of reducing the number of undiagnosed AF patients.
As input, the software takes data from a patient's 12-lead resting ECG (including age and sex). It is only compatible with ECG recordings collected using 'wet' Ag/AgCl electrodes with conductive gel/paste, and using FDA authorized 12-lead resting ECG machines manufactured by GE Medical Systems and Philips Medical Systems with a 500 Hz sampling rate. It checks the format and quality of the input data, analyzes the data via a trained and 'locked' machine-learning model to generate an uncalibrated risk score, converts the model results to a binary output (or reports that the input data are unclassifiable), and evaluates the uncalibrated risk score against pre-set operating points (thresholds) to produce a final result. Uncalibrated risk scores at or above the threshold are returned as 'increased risk' information; uncalibrated risk scores below the threshold are returned as 'no increased risk' information is used to support clinical decision making regarding the need for further referral or diagnostic follow-up. Typical diagnostic follow-up could include ambulatory ECG monitoring to detect previously undiagnosed AF, as described in device labeling. Results should not be used to direct any therapy aqainst AF itself, including anticoagulation therapy.
Tempus ECG-AF does not have a dedicated user interface (UI). Input data comprising ECG tracing metadata (sample count, sample rate, etc.), patient age and patient sex, will be provided to Tempus ECG-AF through standard communication protocols (e.g. AP), file exchange) with other medical systems (e.g., electronic health record systems, hospital information systems, or other medical device data display, transfer, storage, or format-conversion software). Results from Tempus ECG-AF will be returned to users in an equivalent manner.
Here's a breakdown of the acceptance criteria and the study that proves the device meets them, based on the provided text:
1. Table of Acceptance Criteria and Reported Device Performance
The FDA clearance document does not explicitly state "acceptance criteria" in a table format with specific thresholds that the device had to meet. However, it presents the results of the clinical performance validation study, which implicitly represent the device's performance against expected clinical utility. The study endpoints of sensitivity and specificity were "met," indicating they were within an acceptable range for the intended use.
Here's a table based on the provided performance metrics:
| Performance Metric | Reported Device Performance (95% Confidence Interval) |
|---|---|
| Sensitivity | 31% (31% - 37%) |
| Specificity | 92% (91% - 92%) |
| Positive Predictive Value (PPV) | 19% (15% - 23%) |
| Negative Predictive Value (NPV) | 95% (95% - 96%) |
Note: While the document states "Study endpoints of sensitivity and specificity were met," it does not explicitly define what specific numerical thresholds were considered "met" for acceptance.
2. Sample Size Used for the Test Set and Data Provenance
- Test Set Sample Size: 4017 patients, with one ECG analyzed per patient. (Page 6)
- Data Provenance: Retrospective observational cohort study. Data was collected from 3 geographically distinct clinical sites (real-world data). (Page 6)
3. Number of Experts Used to Establish the Ground Truth for the Test Set and Their Qualifications
The document states that the AF status of each patient in the test set was determined based on "duplicate manual chart review" (Page 6). It does not specify the number of experts or their specific qualifications (e.g., "radiologist with 10 years of experience"). This suggests the ground truth was derived from existing medical records interpreted by qualified personnel, but the specific details of those adjudicators are not provided.
4. Adjudication Method for the Test Set
The document mentions "duplicate manual chart review" (Page 6) for establishing the ground truth. This implies at least two reviewers independently reviewed charts to determine AF status. It does not explicitly state a 2+1, 3+1, or other specific adjudication method if there were discrepancies between the duplicate reviews.
5. If a Multi-Reader Multi-Case (MRMC) Comparative Effectiveness Study was Done
No, a Multi-Reader Multi-Case (MRMC) comparative effectiveness study was not done. The study described is a standalone performance validation of the AI algorithm. The document focuses on the algorithm's performance in detecting signs of AF risk rather than how human readers' performance might improve with AI assistance.
6. If a Standalone (i.e., algorithm only without human-in-the-loop performance) was Done
Yes, a standalone performance study was done. The "Summary of Clinical Studies" section describes the evaluation of "the Tempus ECG-AF model" and its observed performance metrics (sensitivity, specificity, PPV, NPV). This is a direct evaluation of the algorithm's performance without the explicit involvement of human readers in the loop as part of the study design for device performance. The device is then intended to provide information to clinicians, but the study described is an algorithm-only evaluation against ground truth.
7. The Type of Ground Truth Used
The ground truth used was based on "a clinical diagnosis of AF in the next 12 months" (Page 6), determined through "duplicate manual chart review" (Page 6) and "sufficient pre- and post-ECG data available to determine that the patient was part of the intended use patient population and to enable at least 1 year of follow-up to determine the presence of a clinical diagnosis of AF." This suggests a combination of medical record review and outcomes data (clinical diagnosis of AF within 12 months based on follow-up).
8. The Sample Size for the Training Set
- Training Set Sample Size: > 1,500,000 ECGs and > 450,000 patients. (Page 6)
9. How the Ground Truth for the Training Set was Established
The document states that the "Tempus ECG-AF model was trained on data from > 1,500,000 ECGs and > 450,000 patients, with 80% of data used for training and 20% of the data used for model tuning." (Page 6)
While it doesn't explicitly detail the methodology for establishing ground truth for the training set, it can be inferred that it involved similar processes to the test set, likely leveraging existing clinical diagnoses of AF from electronic health records or other forms of medical documentation, given the large scale of the dataset. The text does not provide specific details on manual review or expert involvement for the training set's ground truth beyond "data from" ECGs and patients.
N/A