“Would you trust an AI to make your diagnosis if it’s more accurate than the average clinician?”
In a multidisciplinary class discussion comprising students from bioinformatics to MD/PhD programs, diverse perspectives on AI’s role in healthcare emerged. Those immersed in AI development leaned towards accepting AI diagnoses, while clinical-focused students, while open to the idea, remained somewhat hesitant. As an Interdisciplinary Oncology student straddling both technical and clinical realms, I found this discussion intriguing. It highlighted a crucial point — even if AI outperforms the average clinician, there’s a collective reluctance to entirely relinquish clinical decision-making to AI. Why might this be?
Our hesitation wasn’t rooted in distrust of AI systems but rather an acknowledgment of potential pitfalls in their deployment. If you’ve dabbled in data science or machine learning before, you may know that accuracy alone is an inadequate measure of AI performance. Before we accepted the AI diagnoses, we recognized the need for additional information: the data used for training, validation methodology, generalizability, potential biases, predictive values and more.
Transparency and clarity are paramount in AI system deployment and trust — at least for those well-versed in AI systems and it’s validation.
Even with transparency and clarity, for those without the expertise, trusting an AI to make diagnoses might seem akin to relying on an oracle machine for clinical decisions. Consequently, there are no AI systems that are given full clinical decision-making capacity (at least, at the time of writing). So if we, as a society, can’t fully trust AI to make diagnoses or surrender full clinical decision-making capacity, how is AI currently transforming medicine and healthcare?
The Role of Predictive Systems in Clinical Decision Making
While there’s currently no AI system given the capacity to make clinical decisions, predictive AI has found a role in generating new clinical features for clinicians. In the context of cancer care, AI systems are used in predicting a patient’s risk of cancer recurrence or aggressiveness. Examples include Oncotype DX, predicting metastatic risk and chemotherapy benefit for certain breast cancer patients, and Clarity DX, forecasting prostate cancer aggressiveness based on blood PSA levels. Similarly, in my research, I’m developing an image analysis system to provide clinicians with a new dimension of information for clinical decisions.
AI Optimization of Patient-Care Workflows
Predicting clinically relevant information isn’t the only application of AI in healthcare. At the Canadian Cancer Research Conference 2023 (CCRC 2023), Dr. Muhammad Mamdani showcased ways AI projects optimized patient care workflows. AI systems planned nurse shifts, allocated resources efficiently, and optimized patient monitoring workflows before a patient’s condition became critical. Dr. Mamdani’s approach of integrating data science as a core pillar in the hospital setting could potentially alleviate pressure caused by the shortage of trained medical personnel.
AI is currently flourishing in the medical field, and its development seems poised to continue. Researchers in Medical AI are becoming increasingly aware of potential issues, such as lack of generalizability and bias, leading to a movement to standardize AI systems and practices. Education is another significant movement, with organizations pushing forward to spread AI and data science knowledge and awareness across the medical landscape. As we navigate this transformative era, the integration of AI into medicine holds great promise, provided we address and overcome the challenges that lie ahead.