Over 54% of doctors are experiencing burnout and the primary cause is documentation. Doctors spend up to two hours on clerical work for each hour of direct patient care. Documentation burden is the number one contributor to burnout, which costs the US health care system approximately $150 billion each year. Further, the largely free-text notes generated by healthcare providers are difficult to process and re-use for quality improvement and research. There is a significant opportunity to A) reduce clinicians' clerical burden via ambient data collection while B) enabling incorporation of machine learning into routine clinical care by structuring information in real-time from the patient-doctor conversation. We have built a service that saves each clinician hours per day by using remote human scribes, who we make more efficient with machine learning and natural language processing, to automatically document patient visits. The service runs on a laptop or mobile device and converts patient-clinician conversation audio to structured documentation. Our approach has already enabled the service to be priced at half the cost of a traditional in- person scribe. Finalized notes can be saved into all common electronic health record (EHR) systems via the Redox interoperability interface. We will use the human-curated audio-to-structured documentation dataset to further improve preprocessing, automating more of the scribing process, and reducing human effort over time. A time motion study and EHR click log analysis in a clinical setting demonstrated time savings of up to two hours per clinician per day. We have successfully generated documentation for over 1,700 patient visits and our dataset will double in size in 2 months. In Phase I, we address key impediments in the structure and timeliness of current EHR data, both of which can be improved with our novel system when coupled with the real-time generation of machine learning-ready inputs. In order to do so, we propose three aims: 1) extract machine learning-ready inputs from patient-doctor conversation audio; 2) extract machine learning-ready inputs from structured and unstructured EHR data; 3) further increase cost-efficiency and scale up the quality assurance process using machine learning inputs from conversation audio and historical EHR data. For Phase II we will apply our machine learning inputs to predict disease trajectory for multiple disease phenotypes with a focus on type 2 diabetes mellitus and hypertension. These diseases are ideal proofs-of- concept due to their high prevalence, morbidity, and inclusion in incentive programs for closing care gaps. The $6.3B in 2018 Medicare Advantage incentives supports a clear business case for leveraging machine learning to achieve value-based goals ? improving patient care quality and efficiency.!
PredictionHealth has built a service to reduce doctor burnout by automatically documenting patient visits by processing audio from the normal patient-doctor conversation - potentially reducing the $150 billion dollar cost of burnout on the US health system. We propose that integrating and facilitating the application of machine learning in routine clinical care can improve patient care outcomes and efficiency. The primary goal of this Phase I proposal is to further improve the service's cost-efficiency with real-time machine learning predictions based on structured inputs from the patient-doctor conversation and prepare to integrate machine learning as an integral element of each doctor's toolbox.
