One of the key issues with the use of digital technologies in healthcare has been that they need a reliable feed of data to perform as expected and their output needs to be timely so that the clinicians can benefit at the right time. If not, their long-term adoption would be in serious doubt since they will not provide the intended benefits at the right moment. For anyone familiar with the state of data in healthcare, they know that this is not an easy task. Healthcare data is fragmented, often in a chaotic state, and mostly unstructured. Ensuring that the new algorithms fit into the workflow of those using them will be a major hurdle to overcome. The disorganized nature of the data in provider institutions is an issue, not just to develop these models but also operationalizing them in different centers. These technologies need to have frictionless workflow effect or facilitate better workflows for physicians to adopt their use. Fitting into the existing workflow means that they do not result in extra steps for the clinicians or cause delays in their normal workflows.
For example, in radiology where many see as the initial frontier in healthcare for AI algorithms to affect patient care, key barriers will include IT implementation and integrating into the existing legacy information systems. One example of this is AiDoc’s AI solution for diagnosis of intracranial hemorrhage on head CT scans. Our discussions with some of the pilot sites indicated that although the algorithm identified subtle hemorrhages that could potentially be missed by the radiologists, the initial setup and implementation took a long time and was taxing on the IT staff. The multiple interfaces that had to be created between the scanner and the cloud and then the PACS system was not simple. Also, the algorithm had a high false-positive rate, which created extra work and scrutiny for the clinicians and resulted in serious discussions if the long-term use was worth the effort.
At present, the algorithms that have FDA approval are in fact not, for the most part, executable at the frontlines of clinical practice. This is for two reasons: first, these AI innovations by themselves do not change the existing clinical workflows and incentives that enable those workflows. Simply adding AI applications to existing workflows will not change behaviors. Second, most healthcare organizations lack the data infrastructure required to feed the data needed to optimally run these algorithms at the point of care. Also, it is imperative to train the algorithms to (a) “fit” the local population and/or the local practice patterns, a requirement prior to deployment that is rarely highlighted by current AI publications, and (b) interrogate them for bias to guarantee that the algorithms perform consistently across patient cohorts, especially those who may not have been adequately represented in the training data. In addition, ongoing evaluation and re-training must continue after implementation to continue to reflect changes in demographics and practice patterns which inevitably change over time.
