Immunology
April 1, 2022
Article of the Month – April 2022
April 1, 2022Corti
Overview
Corti is an AI company that aims to enhance human capabilities; they believe AI can significantly assist any task. Corti has taken the responsibility to use technology in enhancing health care by assisting physicians, enhancing patient care, and improving healthcare management globally. Corti pursues these goals through its multidisciplinary team of experts in machine learning, design, engineering, mathematics, and neuroscience. Their innovative team collaborates to implement their creative ideas that could change patients outcomes across the globe.
Corti A.I Aspects
Through Corti, physicians can improve patients’ outcomes, decrease risks, and increase their performance across medical systems.
Audia is a software that assists medical personnel through multiple methods, including integrating with audio and video conferencing tools, recording and uploading video and audio consultations, chatbot integration to analyze important textual information, and synchronization with electronic health records.
Through the enhanced patient consultations, Audia ensures that each medical consultation ends with patients receiving the best medical advice possible in the most simplified manner.
Information gathered through Audia helps increase the efficiency of healthcare systems. It allows healthcare organizations to form a better sense of completed consultations, ensuring that providers can improve after every encounter. It also allows the use of information from patient consultations to predict and understand health trends to allow proper allocation of resources.
Through Corti, physicians can improve patients’ outcomes, decrease risks, and increase their performance across medical systems.
Audia is a software that assists medical personnel through multiple methods, including integrating with audio and video conferencing tools, recording and uploading video and audio consultations, chatbot integration to analyze important textual information, and synchronization with electronic health records.
Through the enhanced patient consultations, Audia ensures that each medical consultation ends with patients receiving the best medical advice possible in the most simplified manner.
Information gathered through Audia helps increase the efficiency of healthcare systems. It allows healthcare organizations to form a better sense of completed consultations, ensuring that providers can improve after every encounter. It also allows the use of information from patient consultations to predict and understand health trends to allow proper allocation of resources.
Other advantages of this software include instant access to information from the medical best practices around the world. This allows doctors to act faster and more confidently, while patients benefit from reduced medical error and more accurate decisions. Cloud-based software implies that the physician and the patient do not have to install software to access the system.(1)
Corti IT Solutions
Optimized triage: Audia presents an AI-based patient triage system, which serves to:
- Reduce the risk of liability: The system ensures that the right questions are asked at the right time, thereby lowering the risk of harmful mistakes. When symptoms are elicited, suggestions are made to the physician to improve diagnosis speed and reduce call duration.
- Detect critical incidents: The triaging system identifies critical cases in real-time, such as stroke or cardiac arrest.
- Decrease over-and under-triaging: Information gathered by the triaging system finds patterns that can decrease over- and under-triaging of patients. The physician will get the consultation recorded with Audia then the record is automatically mapped with the AI insights from the file.
- Automatic quality assurance: Audia automatically builds a quality assurance report of each patient’s consultation, reducing administration time.
- Faster Triaging: The system provides customizable protocols that reduce up to 25% of the time spent interviewing patients, creating an accurate but efficient triaging session.
- Leading speech recognition: Audia detects rate-of-speech and understands the true sentiment of the conversation, assisting medics to infer meaning.
Improved performance:
Audia helps optimize the medical staff’s performance by generating automatic feedback based on measuring a physician’s performance against key indicators of success. Audia also supports healthcare facility administrators in developing a better view of the situation, making resolving prominent issues easier.
Recognized health trends:
Resource allocation is imperative to increasing efficiency and improving patient care. Resource allocation requires that resources are more heavily deployed when necessary and saved when not needed. Audia gives managers the tools necessary to prioritize resources according to a given situation.
Covid-19 solutions:
Audia can analyze patient consultations in text, video, or audio formats and detect if there is a high risk of COVID-19. In Seattle, Audia has already learned from more than 100,000 medical interviews about COVID-19 through a partnership with MedicOne. Audia also creates a real-time map showing the location of patients.(2)
Corti, a Promising AI Company
Corti was founded in 2016 in Copenhagen, Denmark. The company raised a total of $3.4 M in funding over three rounds. Their latest funding was received from the grant round on August 1, 2018. Corti is funded by seven investors, including EASME-EU Executive Agency for SME and Nordic Markers.
Corti has been selected as one of Europe’s top 50 startups and chosen as the best AI innovation in healthcare by Venturebeat. It has also won multiple awards, including Europe’s Future Unicorn Award in 2020 for being the most promising startup in becoming one of the next billion-euro companies.(3)
Multimodal Learning for Live Question Tracking in Speech
Speech sequence labeling is formulated by training a multimodal representation from the temporal binding of the audio signal and its automatic transcription. Therefore, a new model to identify questions in real-time in a noisy environment was shown to be more accurate than previous approaches. Being synced to an automatic speech recognition (ASR) output, the model can be applied as a general-purpose speech tagger according to a general medical symptom labeling task.
Identifying and classifying questions in emergency medical telephone calls involved two key aspects: noise and live processing. Noisy environments can interfere with ASR and text labeling, and therefore the multimodal speech labeler MultiQT was proposed. It uses three neural networks that can be connected to diverse temporal input modalities. By taking advantage of the live component, the multimodal alignment and script transcript could be avoided in the algorithms’ training process. A dataset of 525 emergency medical services phone calls was collected. Each question was manually annotated with its start and stop time and assigned with one of 13 predefined question labels and an extra label for any other question that felt outside the original 13 categories.
On average, it took around 30 minutes to annotate each call. For the experiments, they chose the five more frequent question classes. As the question-tracking data was not manually labeled for symptoms, they automatically created standard training and test sets by distributing a list of textual keywords from the ground truth human transcripts onto the audio signal as time stamps. The original list contained more than 40 symptoms.
This study found rigorous questions and medical symptoms in emergency phone calls and proposed an objective approach to real-time sequence labeling in speech.(4)
Out of hospital cardiac arrest (OHCA) is an emergency condition that affects more than 600,000 people a year in the US and Europe. Early recognition is necessary to initiate cardiopulmonary resuscitation before the arrival of emergency services to improve the patient’s survival. However, approximately 25% of all OHCAs are not recognized because it is a challenging condition to be recognized by nonmedical professionals.
A machine learning (ML) system created by Corti was used to recognize OHCA from emergency recording calls to an emergency medical dispatch center in Copenhagen, Denmark, in 2014. To teach the ML system, researchers used a dataset of 108,607 dispatch audio recordings, of which 918 were OHCA calls eligible for analysis. A fraction of the dataset was assigned to training and the rest to validation. Excluded cases included damaged audio files, disconnected calls, cases in which CPR was initiated before the emergency call, and cases where the patient showed signs of obvious death.
Machine Learning as a Supportive Tool to Recognize Cardiac Arrest in Emergency Calls
Out of hospital cardiac arrest (OHCA) is an emergency condition that affects more than 600,000 people a year in the US and Europe. Early recognition is necessary to initiate cardiopulmonary resuscitation before the arrival of emergency services to improve the patient’s survival. However, approximately 25% of all OHCAs are not recognized because it is a challenging condition to be recognized by nonmedical professionals.
A machine learning (ML) system created by Corti was used to recognize OHCA from emergency recording calls to an emergency medical dispatch center in Copenhagen, Denmark, in 2014. To teach the ML system, researchers used a dataset of 108,607 dispatch audio recordings, of which 918 were OHCA calls eligible for analysis. A fraction of the dataset was assigned to training and the rest to validation. Excluded cases included damaged audio files, disconnected calls, cases in which CPR was initiated before the emergency call, and cases where the patient showed signs of obvious death.
Emergency recordings concerning OHCA were identified and selected for a pilot study. The pilot testing was implemented to achieve a comprehensive evaluation of these calls. Time-to-recognition of OHCA was defined as the time between the call being answered and cardiac arrest being established. The results showed that compared with the medical dispatchers, the ML system had higher sensitivity (72.5% vs. 84.1%, p<0.001) with slightly lower specificity (98.8 vs. 97.3%, p<0.001). The system’s positive predictive value was lower than the control group (20.9% vs. 33 %, p<0.001). Time-to-recognition was significantly shorter for the ML framework (median 44 seconds vs. 54 seconds, p<0.001).
The ML framework performed better than the medical dispatchers for identifying OHCA in emergency phone calls. This technology may become a support tool for emergency medical dispatchers and could be applied in other frequent time-critical incidents such as stroke, acute myocardial infarction, or sepsis..(5)
References