Healthcare is one of the industries where the impact of machine learning (ML) technologies can be described as life-saving (no pun intended). From detecting respiratory diseases and malignant neoplasms in medical images to modeling a patient’s response to the proposed treatment strategy, there’s a plethora of diagnostic and therapeutic tasks ML models can facilitate.
Despite the technology’s enormous potential, ML adoption rates in the healthcare industry remain ridiculously low. In the United States, for instance, the percentage of medical organizations using machine learning fluctuates around 1.8%.
In particular, we’re witnessing reluctance to use diagnostic and predictive models, which do not merely facilitate administrative work but are directly involved in the treatment process.
In this article, a team of innovation analysts from Symfa – custom software development company, an EU-based software engineering company, will identify the key challenges of diagnostic and predictive ML model development in healthcare — and provide expert tips for mitigating these challenges.
Top 5 Challenges in Building Diagnostic and Predictive ML Models in Healthcare
Among the factors that prevent healthcare ML solutions from going mainstream are:
• A global shortage of skilled data science and machine learning engineers. According to recent research from SAS, 63% of business executives believe their top healthcare software companies have the largest skill gaps in the AI and ML segments. Meanwhile, only 10% of the world’s leading data science specialists have proven skills to carry out ML development and implementation projects.
• Limited access to quality data. It’s common knowledge that machine learning algorithms are only as good as the training data they’ve consumed. Even in less regulated industries like eCommerce, it might be challenging to scavenge enough training data for implementing a high-performance recommendation engine. While ML engineers utilize plentiful techniques to artificially expand training datasets (e.g., data augmentation, synthetic data generation, transfer learning, etc.), due to privacy regulations, healthcare organizations can seldom access training data that is both representative and inclusive.
• ML model performance issues. There’s no shortage of studies suggesting ML models outperform human doctors in detecting early signs of breast cancer or providing quality healthcare advice to patients. But, as experts rightfully note, such studies are often overhyped. For instance, less than 1% of the 2,000 studies reviewed by a team of researchers from the University of Birmingham met the basic transparency requirements, which would potentially allow the AI systems to be tested on a larger population. When faced with real-world data, ML models often deliver inaccurate predictions or display bias toward individual patients or patient groups, which undermines trust and creates prejudice against AI-assisted diagnosis and medical care.
• Failure to build inherently scalable ML models. Going back to the flawed research issue, many healthcare organizations struggle to replicate the results achieved by prominent AI vendors when implementing custom diagnostic and predictive models. These issues become even more prominent when medical organizations attempt to scale their intelligent systems across use cases and, ultimately, company-wide. While scaling AI is hard across all industries, in healthcare, the problem is exacerbated by stringent regulations, IT system complexity, and a decision-making process involving multiple stakeholders.
• Ethical considerations. In medical ML projects, there is always a tradeoff between implementing diagnostic and predictive models with limited cognitive abilities (i.e., white-box AI) and high-performance algorithms that fail to explain the reasoning behind their predictions (i.e., black-box AI). Until healthcare organizations address the ML explainability issue, neither healthcare professionals nor their patients can be expected to trust medical machine learning and take its recommendations seriously. As of 2023, 60% of US patients express distrust toward artificial intelligence, fearing its bias and questioning the accuracy of algorithmic recommendations.
How to Overcome Diagnostic and Predictive ML Model Challenges in Healthcare?
Most challenges in developing reliable diagnostic and predictive models in healthcare stem from the following factors:
• Lack of technical talent and expertise in ML model engineering
• Limited data for algorithm training and/or its poor quality
• Questionable design decisions made during the ML model training and implementation
• Failure to communicate the benefits of AI to medical personnel and patients
As an organization looking to integrate diagnostic and predictive models into your medical care processes, you could follow this seven-step guide to plan for and successfully address potential ML implementation pitfalls:
1. Start your journey with a thorough assessment of your IT infrastructure, data, and workflows. Hire a skilled business analyst to facilitate meetings with the C-suite and other stakeholders to identify tasks and processes that could benefit from AI.
2. Unless you can hire in-house data science and ML professionals to augment your existing IT team, you should partner with a reliable technology company providing ML consulting and implementation services.
3. Together with your technology partner, audit your data and determine what information you would need to train a custom ML model from the ground up or re-train a ready-made foundation AI model. As your data may come from different sources, such as electronic healthcare records (EHRs), medical imaging systems, laboratory software, and even medical wearable devices, you must take steps to integrate, normalize, and unify it for algorithm training. And don’t forget about meeting region-specific compliance requirements.
4. Be prepared to artificially expand your datasets or generate synthetic data in collaboration with certified physicians — especially if you’re building an ML model for diagnosing and treating rare health conditions.
5. Make sure your ML partner selects the right tech stack for ML model implementation and scaling. This may include a robust cloud platform, instruments for application containerization and container management, and solutions for load balancing. A rule of thumb is to use a mix of proprietary and open-source technologies to avoid vendor lock-in.
6. Start executing your ML strategy with a small proof of concept (PoC) while eyeing the bigger picture. Collaborate with your vendor to validate that the data used for model training is diverse and representative. Consider using performance monitoring tools to timely identify and address issues. In case you’ve opted for a black-box model, tap into advanced techniques like feature importance analysis and Local Interpretable Model-agnostic Explanations (LIME) to add an explanatory layer on top of your model, thus addressing potential trust issues.
7. After deployment, continue to train and adjust your model in collaboration with domain experts. This way, you could kill two birds with one stone. On the one hand, the model would absorb human feedback and improve its performance drastically. On the other, physicians will learn to work with AI, getting to know the tool and appreciating the benefits it has to offer.
In the absence of universal guidelines and regulations for medical ML development and implementation, it’s not easy to tell how many years separate us from a full-blown healthcare AI revolution.
However, the global healthcare AI market is currently growing at a CAGR of 37.5%, which is higher than in the majority of digital-first industries. And this trend suggests that the adoption of machine learning in healthcare is inevitable, despite the challenges it faces.
