The applications of artificial intelligence (AI) in healthcare are accelerating rapidly, with potential applications being demonstrated in various fields of medicine and healthcare. However, there are currently limited software testing tools and strategies for testing AI-based health products. This article then explores the main challenges and limitations of AI-based healthcare software testing and examines the steps needed to improve current testing strategies, potentially transforming technologies from research to clinical practice.
“The exciting promise of artificial intelligence (AI) in healthcare has been widely reported, with potential applications in many areas of medicine.” (Topol, 2019). Healthcare organizations around the world have been well received around the world by AI-powered healthcare systems and fields. Nonetheless, healthcare organizations have yet to realize the potential of AI.
However, an increasing number of products and systems have started to use AI algorithms in clinical practice and clinical systems. The following section will give a brief overview of AI in healthcare.
The growing potential of artificial intelligence in the healthcare sector
A growing body of literature has demonstrated the various applications of AI in healthcare, such as interpreting x-rays, cancer, and other long-term health issues. (Hwang et al. 1,2019).
âAnalyzing the immense volume of data collected from electronic health records (EHRs) holds promise for extracting clinically relevant information and performing diagnostic assessments (Escobar et al. l 2016) as well as to provide real-time risk scores for ICU referral (Liang et al. L 2016), predict in-hospital mortality, risk of readmission, extended length of stay, and discharge diagnoses (Rajkumar et al. 1,2019), predict future deterioration, including acute kidney injury TomaÅ¡ev et al. l, 2019), improve decision-making strategies, including weaning from mechanical ventilation (Prasad et al. L 2019)and management of sepsis, and learning treatment policies from observational data. “
AI has great potential to rapidly improve healthcare outcomes. AI tools and software will shape the future of healthcare delivery with a more individualistic and patient-centered approach.
Software testing in AI and ML
The central element in the development of machine learning and AI algorithms is testing. You can compare this with unit testing of software testing. AI / ML engineers develop an AI algorithm and verify that the training data is doing enough work to accurately classify or regress the data with good generalization without overfitting or underfitting the data. Engineers also use certain validation techniques, which are like test data from software testing.
AI-based software uses algorithms and data together to account for hyperparameter configuration data and associated metadata, which primarily work together to display results. If the validation phase of the algorithm gets wrong parameters that could affect the results we are looking for. To get more accurate results, the engineer must review the algorithms themselves, modify the hyperparameters, and reconstruct the model, perhaps with better training data. This could be compared to the system test, which the tester performed to understand the behaviors of the system.
On the other hand, Al’s engineers could do a lot of work to understand the behavior of algorithms. Sometimes algorithms and models work well; However, when you deploy this in the real world, you can get a lot of errors. Still, we did everything we were supposed to do during the training phase. Our model exceeded expectations, but it does not enter the âinferenceâ phase when the model is operationalized. This means that we have to have a quality assurance approach to dealing with the models in production.
The following section will discuss approaches to QA in AI-based healthcare fields.
AI-ML based healthcare domain software testing approach
A standard healthcare domain test is a process for testing healthcare applications with factors such as security, compliance, cross-dependency with other entities, etc. The tester ensures that the quality, reliability, performance, safety and efficiency of the healthcare application in its place and the software behaves as accepted. Current AI-based tools and software come with algorithms and logic tests, which Al’s engineers have already done.
However, the difficult part for the tester is to test the behavior of the algorithm in the software and the system. QA teams should have domain knowledge and experience with healthcare systems, algorithms and how these two work together etc. Testing an AI-based healthcare solution requires a strategic approach tailored to each tailor-made scenario and the needs of the healthcare system.
Most healthcare algorithms are quite complex and difficult for software testers to predict. The algorithm goes through training and testing sets, creating meaningful data associated with human behaviors. Insufficient or incomplete data set or containing poor quality data can lead to bias in the solution. A system is overtrained to see the same thing or is not trained enough to make an accurate judgment.
Another challenge that testers face when testing AI-powered health systems is the amount of data required to test the system. The restricted data element approach will not provide statistical assurance of the system. This opens the door to yet another challenge for testers as to what kind of skills a tester should have and how he should interact with these systems of this level of complexity.
What skills and approaches are needed for AI-ML based health domain testing?
As we discussed earlier, creating and training the algorithms is a manual, automated process with some of the test items. Testers primarily use boundary testing and double coding to solve most complexity issues. Testers should have some knowledge of the data, and knowledge of algorithms would be an essential skill.
Sometimes, the algorithm used, the volumes of data or the complexity of the solution, testing these systems can be as complex as the solutions themselves. This requires extensive technical and data science expertise on the part of testers, which makes the job of the AI ââtester different from that of any manual or automated tester. However, the agile approach and the interdisciplinary approach to the testing process would help testers better understand algorithms and their functionality.
In conclusion, AI-powered health products will gain momentum day by day. As testers, we need to be prepared to test one of the most complex algorithms and logic, potentially saving lives and protecting people.
Article written by Dr Ali Yildirim, Oxford Health EHR Software Testing Manager