Last week, Derek A. Haas, Chief Executive Officer of healthcare technology and analytics company – Avant-garde Health; Eric C. Makhni, Sports Medicine Orthopedic Surgeon at Henry Ford Health System; Joseph H. Schwab, Associate Professor of Orthopedic Surgery at Harvard Medical School and Chief of Spine Surgery at Massachusetts General Hospital, and John D. Halamka, Executive Director of Health Technology Exploration Center of Beth Israel Lahey Health and the International Healthcare Innovation Professor at Harvard Medical School co-wrote an article published in Harvard Business Review, to address some of the misconceptions around machine learning (ML) and healthcare. 

ML is not about replacing the physicians

There is an ongoing misconception that some medical professionals (especially radiologists) will be eradicated because ML applications are now capable of being trained in the shortest possible time to make sense of a large amount of clinical data and made accurate predictions. However, according to the authors, this will never be the case mostly because the role of physicians are multi-facets and deriving at a diagnosis is just one of the many things they do on a daily basis. 

Apart from that, doctors are also responsible for disease prevention, providing care and cure, remind patients when they are engaging in malpractices which may worsen their conditions and so on. At the moment, ML has a limited capacity to perform some of these responsibilities even though researchers began to look into causal artificial intelligence (AI) and its impact on population health. Nevertheless, even with further development, they will never replicate a human doctor’s ability to render personal concern and care. 

Besides, there is still a need to involve individuals with respective domain knowledge to train an ML algorithm and validate its output to ensure it will perform as intended. Unless all these human elements could be replaced, otherwise, thinking ML could replace physicians is non-practical. 

ML is not about grouping data and data scientists together 

Indeed, ML needs data and data scientists but these are not sole ingredients to a winning recipe. Most ML are trained using data extracted from electronic health records (EHRs) and these data represent patients who had visited and seek treatment from certain medical institutions, not the entire population. So, an ML algorithm runs the risk of not being able to perform as efficiently, if it’s not constantly updated and validated. 

Furthermore, EHRs across different medical institutions tend to be provided by different vendors. Even if they do happen to be using the same ones, there may still be differences in terms of the kind of data being collected, structures of data, field meanings, and so on. As such, additional time and attention are needed to be paid to clean the collected information, before feeding into an ML algorithm. 

Whilst highly skilled data scientists may be able to build complex ML models, they may not necessarily have the domain knowledge to understand and clean the required data. Therefore, the most competent way is a synergy between domain expertise, ensuring that the intended ML algorithm will address the clinical problem. 

ML algorithm is not always adopted and utilized

Earlier, AIMed has touched upon the AI and ML adoption challenges faced by clinical radiology in two successive webinars (you may re-visit them here: one and two). There are many reasons why an ML algorithm is not well-received among medical professionals. It can be an infrastructural challenge: whereby the developed algorithm does not synchronize with the existing system. It can also be a workflow challenge: the ML solution does not incorporate into what physicians or clinicians are doing right now. 

The authors urged readers not to see these misconceptions as setbacks but they should be valuable lessons that those who are interested or are already in the field creating or using ML, should take note of. At the end of the day, there are successful cases. For example, the ML application that “reads” incoming notifications and file them into the correct medical record at the Beth Israel Deaconess Medical Center. It has helped save 120 hours of staff time per month. Thus, ML will work, as long as it is meaningfully applied. 

Author Bio
synthetic gene empathy chinese artificial intelligence data medicine healthcare ai

Hazel Tang

A science writer with data background and an interest in the current affair, culture, and arts; a no-med from an (almost) all-med family. Follow on Twitter.