Beyond the promise of AI in healthcare: How well does the bear dance?
Artificial intelligence (AI) and machine learning (ML) are both terms that have become murky through over-use and misuse. Investors are increasingly sceptical of companies claiming their technology is powered by AI.
They question if, in fact, the promised mechanics are little more than computer-driven automation or merely standard multiple, or even single, linear regressions. Deeper investigation often reveals the software does not begin to approach sophisticated ML or deep-learning (DL) algorithms.
Ultimately, if the machine can ‘See’, ‘Hear’, ‘Read’, ‘Move’ or ‘Reason’, then identifying the respective input, interpreting the data and, ultimately, making independent decisions, will demonstrate the system is true AI. Human health may well be the arena in which AI promises the greatest potential benefit, but, for all its promise, we have yet to see its mass adoption or any substantial improvement in outcomes from this technology.
David Shaywitz, a Harvard-trained physician, lecturer, and former Takeda Ventures partner, has questioned the effects of data science and technology (DST) on health and disease. As he puts it in a Forbes column, we need to move past the “Dancing bear stage” and not just marvel at the mere fact that the bear is dancing at all. The time has come, he argues, to question how well it dances.
Shaywitz interviewed Novartis AG Chief Executive Officer Vas Narasimhan, a physician, former McKinsey consultant and strong proponent of leveraging DST to power drug discovery. Even Narasimhan, when asked about the application of AI to real-world evidence, is sceptical of obtaining statistically significant data when the powers of randomness and blindness inherent to controlled clinical trials are removed. The lack of consensus on appropriate statistical methods to validate AI technologies explains, in part, the reticence of the medical community to adopt new digital therapeutics, biomarkers or tools. Doubts surround the statistical methods to analyse continuous streams of data from wearables, the reproducibility of in silico clinical trials based on algorithms built with unsupervised learning and the appropriate placebo for digital therapeutics that adapt to each patient.
Nevertheless, AI-fueled drug discovery does offer reasons for optimism. An analysis by Boston Consulting Group identified 15 molecules developed with an AI approach that were in clinical trials in Q1, 2022. Although no drug has been approved, the AI-fuelled pipeline has been expanding at an annual rate of nearly 40%. On the medical-imaging front, the premium reimbursement granted to Digital Diagnostics, Viz.ai, Heartflow, Perspectum and RapidAI offers a proxy for adoption by medical societies and a lead indicator for uptake in practice. Still, only the former has secured a permanent classification at the time of writing.
Investors should not be dismissive when companies boast of their AI capabilities. However, for those involved in biotech and MedTech investing, sound data, statistical methodology and reimbursement strategy should be validated carefully to determine whether the promised AI system offers a clinically meaningful benefit to patients and, as a consequence, the ready adoption by healthcare professionals.
