The progress in research due to AI is facilitating Machine Learning (ML) techniques with potential to reduce costs, improve results, and increase patient satisfaction. However, applying ML techniques to create AI models is only a small portion in the overall realization of a clinical solution. Let me illustrate this with an example. Consider two cases – identifying a person from your personal photos vs. identifying a lesion in MR images of the brain. Both of these can be implemented using standard AI techniques but there would be a considerable variance in how quickly they could be deployed in use by the end-user.
So, what does it take to build a real-world AI-enabled clinical solution? First, let me draw attention to the principle challenges that exist today – the acquisition and aggregation of data. In the world of AI, it is often believed that one who has access to data is the king. But, in my view, one who has access to curated, annotated and consent managed data is the emperor of the world!
Unlike the large collections of public datasets that are available for creating general-purpose applications such as ImageNet, CIFAR etc., there is a significant shortage of high-quality medical datasets of specific clinical conditions. A vast majority of medical data is unstructured and dispersed across various systems, which makes data acquisition, extremely complex. Furthermore, region specific data residency and privacy laws also limit the sharing of patient data.
Once the problem of data acquisition is resolved, we are then confronted with the issues of quality and diversity. For example, to develop a lung-abnormality detection application, thousands of CXR images for each of the abnormalities – both in isolation and combination are required. Searching the hospital image archive for specific abnormalities in CXR images is highly error prone due to amount of manual effort required. In addition, for many clinical problems, a single data point is inadequate and needs more elements from the longitudinal record of the patient such as imaging priors, radiology reports, genomic data, lab reports etc.
Before the model creation, the data scientist has to spend a considerable amount of time and effort on the pre-processing steps. For example, to test the condition of the cells in Pathology, pre-processing steps like detection and segmentation of nuclei and glands are required, which themselves demand a highly accurate AI model. Although Deep Learning (DL) models are good at automatically extracting the features, a number of traditional features (like pathomics in pathology) improve the overall accuracy of a classifier. The challenging part is of course to find out which features are relevant and influential on the final prediction model.
Here comes the final and most appealing part of AI – the model development. It is only through close consultation with clinical experts that a data scientist can understand the problem statement for building the AI model. The more adept the data scientist is at choosing the analytical method, the higher will be accuracy of the model. Additionally, choosing a ML vs. DL technique holds significant importance in healthcare since the solution needs to justify the end result.
At Philips, we think of AI-enabled solutions as an extension of human capabilities, rather than a replacement. Any AI-enabled solution needs to demonstrate that it can augment the healthcare provider’s expertise in a meaningful and accurate manner. As the model is trained on a limited dataset, it is critical to be aware of its limitations when tested on datasets from diverse regions.
Despite the progress on the interoperability front, there are very few common representations of clinical pathways in healthcare. An AI model can’t be implemented in a hospital using the same parameters that were used to design it in the lab. Thus, it’s imperative to integrate it seamlessly with the existing clinical workflow of hospitals.
To ease the deployment of AI-enabled solutions in healthcare while keeping the aspects of hosting, performance and security in mind, we decided to build the HealthSuite Insights in Philips based on inputs from the data science and clinical communities.
Here are a few recommendations from us, which can simplify the overall process of creating AI-enabled clinical solutions. Firstly, churn out a cross-functional team of data scientists, clinicians, product managers, application specialists, software engineers, user-experience experts, DevOps and Q-and-R engineers. Then, study the clinical problem statement and role out the limitations of the existing solution. Further, establish clarity on the intended use of the proposed solution and clinical workflow integration. Institute a comprehensive data management strategy including acquisition and curation and acquire data from geographically distributed institutions.
Spend quality time in real-time testing to determine the overall accuracy of the model and focus on matters of privacy including consent management, security and regulatory. To eliminate issues pertaining to scalability, adopt an ML platform that can accelerate the development and deployment phases. Therefore, a comprehensive validation strategy that considers the intended use of the solution, interpretation variabilities, data and workflow diversity can help us confront the challenges and unlock the immense potential of AI in healthcare.
AI is an emerging technology. The healthcare institutions which will invest in open and interoperable standards for data management and clinical workflows today, will pioneer the implementation of AI-enabled solutions tomorrow. This will not only lead to better outcomes for patients but also significantly simplify the overall healthcare system.
In conclusion, I would like to quote an ancient Indian Sanskrit verse that keeps me personally motivated as we undertake this revolutionary journey: