Our team is often asked why machine learning (ML) isn’t more prevalent in healthcare. In this first post of a series on barriers to healthcare ML, we discuss one of the biggest hurdles, data reliability. Most CTOs in the health space are excited about using ML throughout their healthcare systems if they aren’t already. We hear stories of neural nets outperforming seasoned clinicians in prominent research studies, such as predicting mortality based solely off of when lab tests were ordered, or predicting patient characteristics from retinal images. That…
Data Science Blog
The Dangers of Commoditized Machine Learning in Healthcare: 5 Key Differentiators that Lead to Success
Machine learning’s popularity in healthcare is growing thanks to its expanding capabilities in medical image analysis, predictive analytics, and prescriptive analytics for clinical decision support. The machine-learning-as-a-service market is expected to grow to almost $5.4 billion by 2022, with healthcare certainly being one of the industries driving that trend. Even today, many technology companies already deliver machine learning models specific to healthcare. Given how available and attractive this shiny new object is, it’s tempting for health systems to jump on the machine learning bandwagon as they investigate innovative ways to improve operations and reduce costs.
But let the buyer beware. In a certain sense, machine learning in healthcare is being commoditized. The combination of EMR ubiquity, increased computational power, the open source movement, and the rise of cloud providers has made training machine learning models easier than ever. But just because a vendor develops machine learning models and delivers them to a client doesn’t mean it offers a complete package to make machine learning work toward its intended purpose.
Several qualities define machine learning’s effectiveness in a healthcare setting. For example, is the machine learning model accurate? Does it apply to the use case or to the pressing business need of the health system? Does it fit well into the user’s workflow? What machine learning-driven interventions can the user perform to affect the corresponding outcome metric? To answer these questions around properly using machine learning, health systems and payers should consider a few factors that distinguish one machine learning vendor and model from another.
Five Key Differentiators to Consider with Machine Learning in Healthcare
Effective machine learning is the product of more than data, features, and algorithms, and is defined by five key differentiators:
- Vendor’s expertise and exclusive focus on healthcare.
- Machine learning model’s access to extensive data sources.
- Machine learning model’s ease of implementation.
- Machine learning model’s interpretability and buy-in.
- Machine learning model’s conformance with privacy standards.
Differentiator #1: Vendor’s Expertise and Exclusive Focus on Healthcare
A machine learning vendor that’s exclusively focused on healthcare commits its expertise, products, and services to help health systems and payers improve outcomes. This focus implies a dedication to healthcare-specific analytics and decision support technology. Many machine learning vendors are only partially focused on healthcare, also spreading their resources to other industries. Considering the complexities inherent in any healthcare setting, an exclusive and unrelenting focus on the healthcare setting is imperative.
Vendor expertise means knowing the pain points of health systems and payers and the workflows where machine learning can most effectively be leveraged. Many companies that focus on machine learning technology don’t have a healthcare background or subject matter expertise on staff to know, for example, how to build an accurate risk model that’s appropriate for a specific use case, much less how to apply the risk scores that machine learning generates. The vendor should be positioned to not only supply the model, but also provide relevant clinical, financial, and operational healthcare expertise.
It takes experience to understand clinical workflows and their inefficiencies, and to guide clinicians in making connections between risk scores and the actionable decisions that risk scores produce. Product development and professional services teams, with skills honed from years of field experience, intuitively recognize when to apply machine learning to benefit a health system. A vendor with process improvement experience across a broad base of health system and payer partners knows what it takes to change workflows in response to machine-learning generated insights.
Differentiator #2: Machine Learning Model’s Access to Extensive Data Sources
A machine learning model shouldn’t be limited to a single data source, like an EMR. The model needs access to multiple data sources through an analytics platform that can aggregate data from claims, labs, pharmacy, radiology, HIEs, billing, patient satisfaction surveys, multiple EMRs, and more. More data means more accurate models, so clinicians can focus interventions on patients who need them most, ensuring that no patient is accidentally overlooked or unnecessarily treated.
Differentiator #3: Machine Learning Model’s Ease of Implementation
EMRs have generated significant clinical and IT staff fatigue around implementing, learning, and using technology. The prospect of adding another complex technology layer to their workloads could be daunting. If the idea behind machine learning is to create clinical efficiencies, then it must be easy to implement and use. Healthcare experience and technological know-how help expedite machine learning implementation. A best-practice machine learning initiative should work within a health system’s existing IT infrastructure. The majority of practical machine learning models can be trained with less than 16GB of RAM, for example. Installing expensive new servers isn’t necessary and only adds to the cost, time, and energy of implementation.
On the software side, healthcare firms should leverage support from the broader machine learning community that includes experts and others going through a similar implementation process. healthcare.ai is a good example of a community that fosters machine learning model development through education and open source technology tools. Support from healthcare.ai adds value to the implementation process, helps analysts learn data science work, and positions the machine learning vendor as an extension to a health system’s analytics team.
Differentiator #4: Machine Learning Model’s Interpretability and Buy-In
Machine learning must be placed in the right point of the clinical workflow to most effectively identify interventions. Machine learning-based decision support must present all possible interventions and help clinicians make the best choice on a per-patient basis. An ideal model should not only present a risk score, but also provide actionable interpretation of that score. Interpretability is compulsory because clinicians must know why the model is producing certain risk scores and why specific interventions are advised, such that clinicians, not models, can ultimately make the right clinical decisions.
Predicting readmissions is a common use case for machine learning. Many technology companies can deliver a model that’s predictive and generates a risk score, but clinicians need to know what levers (e.g., intervention options, such as scheduling a follow-up appointment, developing a discharge education plan, or scheduling a home health visit) are available within their health system around readmissions or they won’t know how to use the risk score. It’s important to link a risk score to available levers that can directly improve a patient’s outcome.
Healthcare professionals are skeptical of new tools, and for good reason. EMRs have generally been difficult to use. Therefore, it’s critical to get buy-in from the end-user on any machine learning project. This is accomplished throughout the development process. As mentioned above, the model must be interpretable and provide simple, actionable suggestions. If it doesn’t, there’s little chance of the model moving the needle on the associated outcomes metric.
It’s one thing to create a model to make predictions, but it’s more beneficial to know when a use case calls for machine learning, how to tie the appropriate levers to the model, how to effectively alter clinical workflow, how to get buy-in, and how to make the output easy to use.
Differentiator #5: Machine Learning Model’s Conformance with Privacy Standards
Healthcare technology, including machine learning, is bound by certain privacy and security requirements around patient data, particularly when it comes to heeding HIPAA privacy rules. Some machine learning solutions require that data be sent to the machine learning tool’s location, which takes data out of its native, protected environment (i.e., outside of the data warehouse or analytics environment). Moving data to the cloud can be a smart decision as long as the machine learning vendor influencing this decision can meet all of the client’s healthcare analytics needs. A best-practice machine learning solution is flexible enough to deliver the tool to the data, alleviating any privacy and security concerns. Doing the machine learning work in the pre-established analytics environment makes it much easier to deliver a machine learning project on budget and on time.
Consider Every Angle of a Machine Learning Investment
Health systems and payers search for every opportunity to improve clinical and financial efficiencies to help them deliver better care at a lower cost. Machine learning is an emerging opportunity that holds significant promise for fulfilling these goals, but it’s also an investment that is more likely to pay off if health systems ask the right questions about what differentiates one machine learning model—and vendor—from another.