As a Full Stack AI/ML consultancy, focused on helping life science and healthcare clients develop new AI-based products and tools, we spend a lot of time thinking about new AI/ML applications. We think the most valuable emerging or high impact trends in AI/ML for life sciences and healthcare are:
We are seeing a rapid movement for medical device companies to “cloud-enable” their devices and create cloud storage for all the data generated by (or which could be generated by) their devices. Medical device companies are realizing that capturing all of their data, in as high resolution as possible, can be valuable for AI/ML applications (and also understand that, if they don’t collect and use the data, hospitals, competitors or other participants in the health care value chain will do so). The long-term value is primarily to use that data to train AI/ML models to monitor health conditions, provide diagnostics and alerts related to device performance and safety, and to create a differentiated value proposition for the user.
Medical device companies are also learning that if they don’t ALSO incorporate modern concepts like dataOps and MLOps (collectively “XOps”) to allow for collaboration, traceability, model monitoring, and maintenance then they are creating technical debt that will cost time and resources in the future. The FDA action plan for AI/ML starts to lay out some of the standards that we believe will drive the adoption of XOps in medical AI/ML.
Genomic sequencing is increasingly a part of clinical trials and clinical medicine. Large-scale transcriptomic, proteomic, and metabolomic assays are becoming more prevalent. Currently, there are many low-cost, highly utilized public domain tools for bioinformatics but the selection of those tools (and version thereof) tends to be ad-hoc, decided by the individual researcher or bioinformatics specialist. The development of cloud-scale bioinformatics pipelines will unify and standardize the workflow to make R&D on these data sets more efficient. Collaboration (many users analyzing the same data), auditability (being able to replicate and understand how the data was transformed), standardization of applications, and centralized control and monitoring of data input quality requires techniques that are really only practical in cloud-based environments. Huge efficiencies will be gained when every scientist can follow, replicate and modify past data transformations to conduct QA, review feature extraction, and confirm the bioinformatics model used.
AI won’t replace scientists, but it can help them gain new insights from more papers than they could read in a lifetime. NLP tools are increasingly being used to mine biomedical literature to better understand relationships between biological entities to accelerate research and discovery and inform clinical medicine.
No code or low code solutions exist for various ML applications and are becoming available for NLP applications. We’ve delivered several low-code NLP platforms configured to allow researchers to focus on a deep understanding of a specialized technical domain such as a specific disease, area of biology, or a drug’s mechanism of action.
In neurology and neuroscience, the development of digital biomarkers based on behavior captured via video, devices, and other multimodal signals such as EEG and actigraphy, will better inform the diagnosis and treatment of neurobehavioral and neurodevelopmental disorders. Given the difficulty in diagnosis and precise determination of CNS disease progression, there is huge interest in using AI/ML in clinical trials, both for the enrichment of the responding population and measurement of response to treatments over time. Finally, the opportunity to automate data collection over larger populations (potentially in virtual clinical trials) is gaining traction, especially when looking at difficult to diagnose and treat CNS diseases.
One trend we would like to see but has yet to develop is for companies to capture data for digital biomarkers during a trial, much like capturing tissue samples. They may not have the AI/ML application developed at the time of the trial but the opportunity to learn more about the patients by capturing digital biomarker data is uniquely valuable.
Telemedicine is in the early stages of technology development. Many telemedicine services platforms are not much more than secure video conferencing technology at this point. We are seeing significant interest in making these systems more valuable to physicians and patients by incorporating additional AI/ML-driven features, for both the provider and the patient.
Valuable telemedicine-based AI/ML applications are yet to be created but there are a number of intriguing possibilities that companies are looking into: The ability to organize and provide patient information to the provider during the session and to capture information during the session could improve workflow and increase efficiency. The ability to longitudinally reference past cognitive and emotional context over time creates opportunities for better and more efficient care. Chatbot-like capability may also be used to pull in relevant information in an automated manner in a triage situation.
Recording of patient sessions, useful for developing a deep training set, may be impractical and risky from a security and privacy standpoint. That can be overcome - we are developing techniques to generate anonymous features from the streaming video and audio signal, to allow for advanced modeling based on recorded features, without the security risk of storing video or audio content.
Distributed Clinical Trials (DCTs) are a hot topic now. After years of discussion, the pandemic boosted the need for DCT’s and, we believe, accelerated the trend to enroll, engage and monitor patients with fewer on-site visits. A number of companies providing basic workflow and communication type services are growing rapidly and we believe that the fact that more of the data and communication will now be recorded electronically will lead to huge AI/ML opportunities.
For instance, attrition of clinical trial participants remains a major concern and leads to excess cost and failed studies. The use of ML models for engagement and behavior has been shown to improve engagement in non-healthcare consumer industries and the same approaches are coming to clinical trial recruiting and patient management. Being able to tell what patients are at risk of dropping out and understanding how to recruit more effectively is a huge competitive advantage for sites, CROs, and trial sponsors when you consider the benefit of getting a trial completed faster.
Also, data from participants can now be continuous, or at least more complete than one data point per site visit, and that will lead to extensive opportunities around digital and biological biomarker discovery. These biomarkers can be used to drive decisions in subsequent trials, like identification of responders for clinical trial enrichment. The bottom line is that the emergence of DCT’s will result in the availability of more data that can be mined for value, perhaps in ways that are not yet appreciated.
AI/ML-driven outcome predictions, based on increasing access to real-world healthcare data, will help healthcare payers prioritize resources and add more automation around burdensome processes such as pre-authorization and payment. Increasingly, NLP is being used to mine clinical documentation and claims in order to train AI/ML applications to predict behaviors, outcomes, and costs to drive actions (early settlement of claims or optimal therapies, for instance). As the move to value-based care continues, more approaches to predict HEDIS scores and other quality metrics will be valuable to optimize care delivery and interventions.
