Artificial Intelligence (AI) and Machine Learning (ML) have been buzzwords for a while now, 2017 was certainly a year with a lot of hype and investment. However, the substance of, “What does AI actually do?” has yet to really be actualized by most business users. Part of the reason for that comes back to the business models that emerged as AI and ML took off over the last five years. These business models represent the first generation of AI investments and had some common themes:
The gaps created by these types of business models have created new opportunities for “second wave” AI companies, or AI 2.0. This new set of companies, of which Precision Health AI is one, hastwo significant differences from AI 1.0 companies. First, AI 2.0 companies incorporate domain-specific knowledge into theirplatforms that reduce customization needs on deployment. Second, in an AI 2.0 platform, the AI model is already trained, optimized and capable of learning versus an untrained toolbox capable of anything but good at nothing.
The exciting and disruptive thing about these AI 2.0 businesses is that they fundamentally represent human intelligencemore accurately. They incorporate context, domain expertise and unique relevant data. They also have the potential to be extremely disruptive not only to AI 1.0 companies, but also to existing B2B businesses that represent the old guard of tech infrastructure.
In an excellent piece written in April 2017, JerryChen, partner at Greylock, describes the existing business technology layers of “System of Engagement”(Apps, EMR, etc.) and “System of Record”(ERP, CRM, Cloud, etc.). As both of these commoditize, there is the potential for the development of new economic moats around the “System of Intelligence” or domain-specific AI trained to operate in a specific business context. This means that both the “System of Engagement” and”System of Record” will increasingly look to the “System of Intelligence” for differentiation and to drive user eyeballs and storage /compute fees.
Just as Netflix’s impact on user engagement has led to a gold rush in video and content production, a similar demand from business users is driving the production of pre-trained AI modules or AI apps to run within existing “Systems of Engagement.”
The structural demand for large “Systems of Engagement” dictates that the majority of financial resources within these companies are directed to satisfying user experience demands. This means that on the whole there has been a minimal investment in developing domain-specific AI themselves. Even where there is internal AI driven R&D, most “Engagement” companies are realizing that the user demand for “Intelligence” will outpace their own capability and they may be better off serving as a channel for innovation in AI-driven intelligence developed and trained elsewhere. Epic, the Electronic Medical Record with leading share in the US, has already begun to explore this model in the marketplace seeking third parties and even customers to develop additional modules to be presented through theirsystem. This seems wise, for if the goal is retention of existing customers, becoming the channel that presents new and innovative content to clients will inherently drive customer satisfaction and retention. While the business model around AI content is still emerging, its importance to existing technology infrastructure will likely drive adoption and partnership.
Over the last year, the team at Precision Health AI has focused on two specific questions:
Leveraging our proprietary deep longitudinal data set, including our partnership with ASCO CancerLinQ,our team of data science, oncology, and AI experts hasbuilt over 60 specific pre-trained AI modules that do specific work in oncology. These productized modules are available both through our Eureka Health Platform and through API and can be run on our data or other local data sources. They can operate as an “Intel Inside” of “Systems of Engagement” or as “Premium AI” or “Power AI” toolkits on top of existing AI platforms like Google Cloud Platform or Microsoft Azure. More importantly, they actually do something useful, and they do it transparently.
Solving real clinical problems: predicting Chemotherapy Induced Neutropenia(CIN)
One area of major challenge in oncology therapy is figuring out which patients can tolerate sometimes toxic treatments. Physicians know what risk factors to look for but there is some guesswork to predicting these types of events. For pharmaceutical companies, making sure patients get the right therapy for them is critical not only for avoiding side effects, but also maximizing efficacy.
CIN occurs in approximately 20 to 25 percent of patients who receive therapy. In these patients, the immune system is weakened and the chance for complicated infection or even death rise by 20 to 25 percent. Imagine having severe cancer and actually dying of an infection that you caught because of a drug side effect.
One of the modules in our Eureka Health Oncology AI Platform is called Onco-Adverse. This module specifically predicts CIN in Lung and Breast Cancer patients. Trained on our deep clinical dataset, the module has an AUC/ROC which exceeds published levels by 0.25-0.4 (extremely accurate) and perhaps, more importantly, leverages a neural net that has thousands of predictive elements. This type of module has the potential to recommend additional testing to clinicians or simply better profile patient cohorts for research and clinical trial selection.
Solving real data problems – and data analysis headaches – by predicting cancer stage For business users who have worked with oncology data, the types of problems that exist are almost too many to count. At a fundamental level, it is difficult to even tell from the data what type of patient took which therapy. It is interesting to know that a breast cancer patient took “drug X”, but what you specificallywant to know is that a Stage 3A patient took that drug. NCCN Staging in cancer dictates treatment recommendation. Most claims data lack this information entirely, and EMR data may contain it but usually the data is inconsistent at best. This missing data requires data analytics teams to do a ton of manual curation of data to get information like stage. Our Onco-Stage module has the ability to predict stage of disease for several oncology indications. While this could play a role in actual diagnosis and documentationlonger term, today it is extremely useful as a data workflow tool for those in healthcare who work with oncology data. By making data more actionablefor business users, the Eureka Health Oncology AI platform enables more rapid insights from data and reduces manual curation time.
Accompanying the hype around AI has been the concern about replacing human tasks and changing the user interface with technology. To be crystal clear, nobody will have their cancer treated by an AI robot anytime soon. The physician-patient interface serves multiple roles, particularly in oncology,where thecritical nuance in the care of patients doesn’t show up in the data. However, any oncologist I’ve met would love more real-time intelligence about patients withReal World Evidence. Physiciansare certainly capable ofinterpreting thatintelligence, butthey do not currently have the opportunity because they are relying solely on underpowered control trials as a source of evidence for treatment selection. Further, and maybe more importantly, everyphysician is dramatically overburdened by process, care coordination, documentation and information overload. AI promises to positively impact the daily lives of physicians and patients by streamlining repeatable tasks. Because of AI, optimizing the process of care so that physicians can use their talents on the truly challenging decisions in oncology is within reach.