Tyto Insights for Crackles Detection

Tyto Insights for Crackles Detection is an AI-driven, over-the-counter software system that analyzes lung sounds captured by a compatible digital stethoscope to identify suspected crackles in adults and children two years and older. Its methodology leverages a cloud-based Convolutional Recurrent Neural Network (CRNN) which processes acoustic signals and has demonstrated high accuracy (AUC 0.97) and non-inferiority compared to clinical readers in detecting this specific adventitious lung sound.

The Tyto Insights for Crackles Detection is an over-the-counter (OTC) software system designed to support the evaluation of lung sounds in adults and children aged two years and older. Its primary purpose is to automatically analyze acoustic signals recorded by a compatible Tyto Stethoscope and identify instances where crackles, a specific abnormal lung sound, are suspected. The system operates using artificial intelligence, specifically a Convolutional Recurrent Neural Network (CRNN) model trained to detect crackle events in lung sound recordings. Audio data acquired by the Tyto Stethoscope, either by a healthcare professional or a lay user, is processed through a cloud-based infrastructure comprising an Application Server, an Algorithm Server (which houses the AI model), and a Web Server that delivers the analysis results (positive, negative, or unable to analyze) to both patient and clinician interfaces. Performance evaluation of the Tyto Insights for Crackles Detection was conducted using a retrospective dataset of 446 lung sound recordings from 445 patients. The AI demonstrated strong stand-alone accuracy with a sensitivity of 0.72 and a specificity of 0.99 for detecting crackles, benchmarked against a ground truth established by a consensus of three pulmonologists. In clinical performance comparisons, the device's ability to detect crackles, measured by the Area Under the Curve (AUC), was 0.97, compared to 0.77 for clinical readers (non-pulmonologists). This performance met the primary endpoint of establishing non-inferiority, with results also demonstrating a notable advantage for the device, and it showed superior consistency with perfect repeatability (kappa 1.0) compared to the readers' moderate repeatability (kappa 0.42). The system also includes a Predetermined Change Control Plan, allowing for future AI model improvements, including re-training with additional data and expanding compatibility with other stethoscopes, under defined validation protocols.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. How does the AI algorithm differentiate between fine and coarse crackles, and does this differentiation impact its reported sensitivity, specificity, or clinical utility?; Given the OTC availability and potential for lay-user acquisition of lung sounds, what is the observed impact of sub-optimal recording quality (e.g., incorrect stethoscope placement, excessive ambient noise) on the algorithm's performance and its 'unable to analyze' rate in real-world usage?; The device's AUC was 0.97 against a ground truth established by three pulmonologists. What was the inter-rater reliability (e.g., Fleiss' Kappa) among these three pulmonologists, and how does it compare to the device's consistency?; The Predetermined Change Control Plan allows for re-training the ML model. What are the specific strategies to ensure that new training data adequately represents diverse demographic groups and a wide range of crackle presentations to mitigate potential biases and improve generalizability?; How does the system's performance for crackle detection vary in patients with comorbid respiratory conditions that may present with multiple adventitious lung sounds (e.g., wheezes, rhonchi) simultaneously, and are there plans to expand detection capabilities to other lung sounds?