An introduction to the healthcare AI lifecycle

AI holds great promise for advancing the practice of medicine. Modern evidence-based medicine (EBM) practice relies on carefully collected scientific evidence, typically from research studies, to make clinical decisions regarding patient care and depends on synthesizing copious amounts of information across large populations of patients.

Machine learning (ML) is a subdomain of AI that provides techniques to build data-driven prediction models to fulfill the goals of EBM. ML algorithms enable computers to learn and improve from data without being explicitly programmed. Much of the recent AI hype has been driven by a type of ML known as Deep Learning. Deep learning uses large quantities of data to train large (deep) neural networks. These algorithms can analyze text, images, and other data types at near-human performance levels.

Complex medical information synthesis tasks may be aided using AI models. For example, AI tools have been developed and implemented to detect rare and serious infections in hospitalized patients. Over 500 Food and Drug Administration (FDA) certified AI systems support physicians in various tasks, ranging from electrocardiogram analysis to mammogram breast cancer detection. Health systems and health information technology (HIT) vendors are also developing and implementing AI systems that do not require FDA certification. These systems are meant to inform physicians by providing risk estimates that can be incorporated into medical decision-making.

These systems are prime examples of healthcare AI, tools used by patients, clinicians, and health systems to improve or inform health, clinical decision-making, or operations. As alluded to above, many different types of healthcare AI systems exist. Despite different use cases, these systems share similarities in making and using them. These similarities directly result from the requirements necessary to ensure safety and efficacy, regardless of the use case. I refer to all of these steps as the healthcare AI lifecycle.

There are two major phases to the lifecycle:

• Development – making AI tools (known as models)
• Implementation – using the AI tools. I would argue that like all software development, making healthcare AI is a continuous journey, so it doesn’t end once a model is in use.

Healthcare AI development & implementation lifecycle. Development is the creation of models. Implementation is the integration of models into clinical care.

The development phase involves the process of creating an AI model. Data access is one of the first issues developers must contend with during this phase. Having obtained data, model developers may realize that healthcare data, like healthcare itself, is complicated.

Processing and transforming data for AI model development requires a unique mix of clinical and technical expertise. After being developed, models must be internally and externally validated (in some settings) to assess if they will benefit patients, physicians, or healthcare systems. External validation may be particularly challenging due to data-sharing restrictions.

Implementation is integrating and utilizing an AI model into clinical care. The implementation process may begin once a model is validated. Implementation raises a host of issues that are rarely confronted during model development. The technical work needed to implement models often requires cobbling together disparate HIT systems, such as databases, web services, and electronic health record (EHR) interfaces.

Additionally, implementation requires special attention to human factors and systems design. Human factors are considerations for how humans interact with technology and the environment, including user interfaces, workflows, and ergonomics – these models are not used in a vacuum.

Finally, there is the issue of monitoring and maintaining these AI systems. As healthcare systems change, AI systems may experience performance degradation due to patient populations or medical practice changes. Thus, developers may need to update their models over time. Despite their promise, successfully developing, implementing, and periodically updating AI models for healthcare is a challenging engineering task.