Search Results

The Eko Foundation Analysis Software with Transformers (EFAST) is intended to support healthcare professionals in the evaluation of patients' heart sounds and electrocardiograms (ECGs). The software calculates heart rate, detects the presence of murmurs associated with Structural Heart Disease, and determines murmur timing including the timing of S1, S2 heart sounds. When ECG is available, the software detects the presence of atrial fibrillation and normal sinus rhythm.

The software is intended for use by or under the supervision of a healthcare professional and is not to be used as a sole means of diagnosis. The interpretations of heart sounds and ECG offered by the software are designed to support and not replace the healthcare professional's clinical judgment. The murmur detection is intended for use on both pediatric and adult patients, while the detection of atrial fibrillation is intended for use on adults (> 18 years).

Eko Foundation Analysis Software with Transformers (EFAST) is a cloud-based Software as a Medical Device (SaMD) intended to provide clinical decision support to healthcare professionals (HCP) in the evaluation of patients' heart sounds (phonocardiogram, PCG) and electrocardiograms (ECGs). The software employs signal processing techniques and machine learning (Deep Neural Networks) to perform simultaneous analysis of recorded heart sounds and ECG data (when available), and identify the presence of murmurs associated with Structural Heart Disease, and determine murmur timing including the timing of S1, S2 heart sounds. When ECG is available, the software also detects the presence of atrial fibrillation and sinus rhythm.
The software does not identify other arrhythmias. In addition, it calculates the heart rate of the patient.

The EFAST Software accepts input consisting of heart sounds and ECG waveforms recorded using Eko Digital Stethoscopes that are saved in .WAV file format and stored in the Eko Cloud, which utilizes the Amazon Web Services (AWS) Simple Storage Service (S3) for data storage and for easy communication with the EFAST analysis server. When an API Request is sent to the EFAST API, the corresponding recordings (PCG and/or ECG) are accessed by the EFAST software and they are analyzed by the EFAST software components. After analysis, the EFAST software returns a JSON format data structure with the algorithm results, which is saved back to the Eko Cloud and can be displayed on EFAST API integrated target software or devices (e.g. user-facing apps).

The EFAST Software consists of the following algorithm components:

• Heart Sound Signal Quality Classifier: Evaluates whether an audio recording contains heart sounds of sufficient quality for further clinical analysis
• Structural Murmur Classifier: Determines whether a good-quality heart sound recording contains a structural murmur or not
• Heart Sound Timing & Murmur Timing: Determines the timing of S1 and S2 heart sounds and determines whether a structural murmur occurs during the systolic or diastolic phase
• ECG Rhythm Classifier: Analyzes ECG signals and categorizes them into one of four classifications: Sinus Rhythm, Atrial Fibrillation, Unspecified Rhythm, and Poor ECG Signal
• Heart Rate Calculation: Signal processing algorithm that utilizes ECG and PCG to calculate heart rate value in beats per minute

Here's a detailed breakdown of the acceptance criteria and the studies proving the device's performance, based on the provided FDA 510(k) clearance letter:

Acceptance Criteria and Reported Device Performance

Note on Acceptance Criteria: The document does not explicitly state numerical "acceptance criteria" but rather presents the device's performance results in comparison to a predicate device, implying that the demonstrated performance is considered acceptable for substantial equivalence. The predicate device's performance often sets an implicit benchmark.

Study Details: Structural Murmur Classification

2. Sample size used for the test set and the data provenance:

• Test set sample size: 2,460 unique heart sound recordings (from 615 unique subjects). 259 recordings were excluded due to lack of cardiologist agreement.
• Data provenance: Retrospective study of de-identified data. Three sites within the US (2 sites) and Canada (1 site).

3. Number of experts used to establish the ground truth for the test set and the qualifications of those experts:

• Number of experts: "Multiple cardiologists" annotated the recordings.
• Qualifications: "Cardiologists." No further specific qualifications (e.g., years of experience) are provided in this document.

4. Adjudication method for the test set:

• For 2,460 initial recordings, "multiple cardiologists" annotated them.
• For 259 of these recordings, the cardiologists "could not come to agreement," leading to their exclusion from the analysis. This implies some form of consensus or agreement-based adjudication, where a lack of agreement led to exclusion rather than a specific tie-breaking rule (e.g., 2+1, 3+1).

5. If a multi reader multi case (MRMC) comparative effectiveness study was done, If so, what was the effect size of how much human readers improve with AI vs without AI assistance:

• No MRMC comparative effectiveness study involving human readers with vs. without AI assistance is described in this section for structural murmur classification. The comparison is between the EFAST algorithm's standalone performance and the predicate device's standalone performance ("Algorithm Performance Comparison to Predicate").

6. If a standalone (i.e. algorithm only without human-in-the-loop performance) was done:

• Yes, the presented performance metrics (Sensitivity, Specificity) are for the EFAST algorithm operating in a standalone capacity ("EFAST Structural Murmur Classification Performance" and "Algorithm Performance Comparison to Predicate").

7. The type of ground truth used:

• Ground truth was established by "pairing the cardiologist annotations with gold standard echocardiogram." This indicates a hybrid ground truth, combining expert interpretation with objective clinical imaging (echocardiogram).

8. The sample size for the training set:

• Not specified in this document. The document only mentions validation on "proprietary datasets."

9. How the ground truth for the training set was established:

• Not specified in this document.

Study Details: Heart Sound Timing (S1/S2 Detection)

2. Sample size used for the test set and the data provenance:

• Test set sample size: 96 heart sound recordings.
• Data provenance: Retrospective study. Data collected using Eko CORE or Eko CORE 500 Digital Stethoscopes. Country of origin not explicitly stated for this subset, but the broader structural murmur study involved US and Canada.

3. Number of experts used to establish the ground truth for the test set and the qualifications of those experts:

• Number of experts: "Expert cardiologists."
• Qualifications: "Cardiologists." No further specific qualifications are provided.

4. Adjudication method for the test set:

• Not explicitly stated, but "annotated by expert cardiologists" implies expert consensus for these 96 recordings, as no exclusions due to disagreement are mentioned. "Cardiologists were blinded to all subject demographic data and echocardiogram findings during annotation."

5. If a multi reader multi case (MRMC) comparative effectiveness study was done, If so, what was the effect size of how much human readers improve with AI vs without AI assistance:

• No MRMC comparative effectiveness study described.

6. If a standalone (i.e. algorithm only without human-in-the-loop performance) was done:

• Yes, the presented performance is for the EFAST algorithm in a standalone capacity ("EFAST Heart Sound Timing Performance").

7. The type of ground truth used:

• Expert cardiologist annotations of "timing of the S1 and S2 sounds."

8. The sample size for the training set:

• Not specified.

9. How the ground truth for the training set was established:

• Not specified.

Study Details: ECG Rhythm Classification (Atrial Fibrillation Detection)

2. Sample size used for the test set and the data provenance:

• Test set sample size: 1,256 ECG recordings (from 314 unique subjects). 97 recordings were excluded due to lack of expert agreement.
• Data provenance: Retrospective study of de-identified data. Two sites within the US.

3. Number of experts used to establish the ground truth for the test set and the qualifications of those experts:

• Number of experts: "Multiple experts."
• Qualifications: "Electrophysiologists, certified cardiographic technicians." No further specific qualifications (e.g., years of experience) are provided.

4. Adjudication method for the test set:

• "Multiple experts" annotated the recordings.
• For 97 recordings, "the experts did not come to an agreement," leading to their exclusion. Similar to the structural murmur study, this indicates an agreement-based adjudication where disagreement led to exclusion.

5. If a multi reader multi case (MRMC) comparative effectiveness study was done, If so, what was the effect size of how much human readers improve with AI vs without AI assistance:

• No MRMC comparative effectiveness study involving human readers with vs. without AI assistance is described. The comparison is between the EFAST algorithm's standalone performance and the predicate device's standalone performance ("Algorithm Performance Comparison to Predicate").

6. If a standalone (i.e. algorithm only without human-in-the-loop performance) was done:

• Yes, the presented performance metrics (Overall Sensitivity, Overall Specificity) are for the EFAST algorithm operating in a standalone capacity ("EFAST Atrial Fibrillation Detection Performance" and "Algorithm Performance Comparison to Predicate").

7. The type of ground truth used:

• Expert annotations related to "quality and the presence of atrial fibrillation."

8. The sample size for the training set:

• Not specified.

9. How the ground truth for the training set was established:

• Not specified.

Study Details: Heart Rate Calculation

2. Sample size used for the test set and the data provenance:

• Not explicitly stated, but the performance is presented as a summary ("found to be less than 5%").

3. Number of experts used to establish the ground truth for the test set and the qualifications of those experts:

• Ground truth method not explicitly detailed in the summary, but typically for heart rate, it involves reference devices or manual counting from a known good signal.

4. Adjudication method for the test set:

• Not explicitly stated.

5. If a multi reader multi case (MRMC) comparative effectiveness study was done, If so, what was the effect size of how much human readers improve with AI vs without AI assistance:

• No MRMC comparative effectiveness study described.

6. If a standalone (i.e. algorithm only without human-in-the-loop performance) was done:

• Yes, the performance is for the algorithm itself.

7. The type of ground truth used:

• Implied to be a reference standard for heart rate measurement.

8. The sample size for the training set:

• Not specified.

9. How the ground truth for the training set was established:

• Not specified.

Ask a specific question about this device