Challenges to applying ML in healthcare

A number of trends have paved the way for increasing adoption of machine learning (ML) in healthcare.

• We’re capturing more volume and types of health data than ever.
• Software for ML are evolving fast.
• Hardware advances have made the computing power cheaper, more agile, modular and scaleable than ever.

However, despite these significant advances, adoption of ML in healthcare remains challenging.

I think some of the major hurdles are:

Data quality

Healthcare data vary significantly in completeness, consistency, accuracy. These differences arise from incentives that drive healthcare providers, the level of training they receive, the culture and complexity of different institutions or places of service.

As a result, errors and gaps are common. Lots of data cleaning (often manual) has to take place before any proper analysis can be done.

As such, ML that automatically ingests, processes data would inherit all these data quality issues, making the models less sensitive to “true” patterns, and are often led down spurious, anomalous paths.

Working with multiple data sets

Healthcare data traditionally exists in different silos. While some commonality exists, different sources usually have materially different data structures. Any analytic effort requires gathering, restructuring, and combining various datasets. This process can be complex, manual and subject to change. Whenever institutions change IT systems, enormous challenges are created in stitching together patient data over time. All these issues make automated ML processes difficult.

Such deviations in data structure also makes scaling up the deployment of ML algorithms challenging.

Subscribe to my newsletter to receive future posts.

Medical complexity

Medicine is complex. From thousands of disease etiologies, to physicians’ clinical judgments, to numerous treatment modalities, to regional practice differences, there can be endless permutations of patient profile and service delivery. ML algorithms can struggle to home in on meaningful patterns given these complexities.

Consistency over time

One of the core tenets of ML is that the algorithms can recalibrate more data arrive. While this is a strength, the dynamic nature of patient profiling can make it difficult to compare clinical outcomes. For example, we may want to identify patients to gauge whether a specific drug is helpful in addressing a medical condition, akin to how clinical trials work. If the patient profiling algorithm changes over time, the comparison between control vs intervention groups may become inconsistent.

Deployment challenges

Assuming we manage to refine the ML methodology to arrive at an effective application, what you may find often is that the recommendations may not be as well received by doctors as you anticipated. For a variety of reasons, the doctors may not find the findings useful. There may be salient aspects of the patient that the ML algorithm could not detect. The doctors may not care about yet another practice alert. (Read this post about practical considerations in ML)

Time, capital investment vs ROI

All these issues make implementation of ML in healthcare challenging. From what I have seen, ML has very high efficacy in specific use cases, such as imaging processing, where the input data is generally uniform, and the traditional human workflow has been manual. Wider applications of ML have had limited success, making justification of the time, resource investment tenuous.

Some ways forward

Subscribe to stay tuned. In the next post, I’ll offer some recommendations to successful implementation of ML in healthcare.