Introduction to Machine Learning In Healthcare

Machine learning applications in healthcare was a great hit with the NYC audience. At least 130 enthusiastic attendees joined the Bots and AI Meetup on December 10th, with the crowd extending far to the back of the room.

Lucy He of Flatiron Health kicked off the night with an examination of machine learning’s impact in medical study cohort selection. A recurring theme was machine learning augmenting the work of human and focusing their efforts. Electronic Medical Records (EMR) are unstructured and can often be quite inscrutable, free-flowing and resistant to complete automation so humans are often involved in tagging and extracting data. That time is expensive so machine learning can serve to make predictions on what patient data is likely to be a fit for a study and prioritize and reduce the workload of the human curators.

Another key aspect of effective cohort studies involves ensuring that the cohort isn’t biased reducing its effectiveness. One bias measurement technique involves identifying and comparing the distributions of clinically relevant variables in an ML generated cohort compared to the reference standard.

Videos of the complete talks by both Flatiron Health and TalkSpace can be viewed on YouTube:

Michael Frank, Director of Strategy at Pfizer, provided a number of use cases improving drug design R&D efficiency. AI work with Generative Adversarial Networks (GANs) can predict and propose more effective molecule shapes. Machine learning in healthcare can also predict and improve potency and yield of molecular compounds. Lastly, ML algorithms can ingest data and draw parallels and conclusions across disparate data sets and research papers on indications, disease pathways, efficacy, and toxicity of individual therapies as well as interactions between therapies.

Image recognition, often deep learning, for identifying pathology such as tumors in medical imaging is making large strides. Another interesting avenue that is emerging is using wearables that collect data and predict impairment or risk. Wearers can be told to proactively visit a practitioner before a worse outcome manifests, a stroke risk for example.

Augmenting business decision making is also a powerful applied AI capability. Trends and momentum can be highlighted by machine learning algorithms exploring industry areas. They can also serve to identify white space where a crowded market still has opportunities for entry.

Finally, Nick Lamm of Talkspace brought the evening back to the chatbot roots of Bots & AI. Of particular interest to healthcare audiences, Nick explained some aspects of vendor and tool selection for HIPAA-BAA compliance and some nuances with cloud services. While Amazon can be used as a cloud platform, not every service available there is compliant. Talkspace is a fan of Rasa Core for keeping local chat data ownership.

Talkspace is experimenting with NLP but their current early release is using smart and guided dialogs. Creating a bot with crafted dialog choices and personalized context memory is often a more successful technique to establish user journey effectiveness prior to adding intent misclassification risk added by free text analysis with NLP.

Marketing and Psychology are not an unusual combination. Motivational interviewing is not just an effective psychotherapy technique but also a powerful sales/marketing technique that can be leveraged even in chatbots to help build trust/rapport and ultimately gain customers.

Machine Learning in Healthcare is such a large space of opportunities and the audience appreciated being able to take in three perspectives. After the event, we were overwhelmed with the number of excited audience members eager for further healthcare themed content as well as interest in speaking.