A Spotlight on Artificial Intelligence from the 2020 Precision Medicine World Conference
This January in San Jose, CA, the Precision Medicine World Conference (PMWC) 2020 took place and included presentations from top industry leaders, researchers, innovators and medical professionals from the biotechnology and healthcare industry.
The event offered seven very interesting tracks, one of which focused on Artificial Intelligence (AI) and Data Sciences.
This track proved to be very thought-provoking, offering case studies where AI and Machine Learning (ML) technology were used to successfully predict clinical trial outcomes, provide bedside decision support in the form of alerts and patient prioritization, and assist with diagnostics by detecting abnormalities in medical images, among other examples.
The results of the case studies presented at PMWC proved compelling from both a patient and financial perspective, demonstrating significant cost reduction (e.g. missed appointments in the UK cost 1B annually), improved operational efficiency (e.g. device alerts in real-time vs. hours in some cases where clinician availability is limited), better clinical outcomes (e.g. predictions leading to increased follow up visits for at-risk patients) and faster drug approval times.
Generally speaking, an AI/ML model to support clinical decision-making is designed to receive and aggregate data from various sources (e.g. CROs, IoT, clinical trial systems, clinician feedback, etc.), process data through algorithms and ultimately generate meaningful predictions. These predictions help clinicians make data-driven decisions for patients while continuously learning from clinician feedback (i.e. to improve accuracy over time. Although several models such as these returned relatively accurate predictions on the clinical deterioration of patients, several future challenges were highlighted including limiting human bias in the design, meeting regulatory expectations and quality/ integrity of real-world data.
Real-world data has been used successfully in the past for clinical trial recruitment, trial design, and marketing but as we move into heavier reliance on AI/ML for clinical assertions, data validity and model accuracy are critical. When AI/ML is relied on to make predictions or draw conclusions in a regulated environment, risk-based controls should be considered and documented to assure that data used for training, testing and ingestion streams (IoT, EHR, images, etc.) are from controlled, valid sources and that regulatory expectations have been met. Several presenters emphasized that that machine learning models are only as good as the data that you feed it.
A key ‘takeaway’ message from the AI/ML track at the PMWC 2020 was that the industry will need to standardize data and improve the interoperability of aggregated data from disparate sources to fully enable AI ML capabilities. To quote one presenter, “It all starts with good data”. ‘Good’ data results from processes and controls well ‘upstream’ of the aggregated data feeding the AI/ ML model including data integrity mitigations, qualified instrumentation, trained personnel, and defined workflows. Incredible advances have been made in building AI/ML models recently as was made clear at PMWC 2020, but we still have significant challenges ahead concerning the quality of data flowing from various sources and in different formats.