Artificial intelligence (AI) adoption in the healthcare sector has grown significantly over the last decade, with no signs of slowing down. According to a 2021 report from HIMSS, 62% of clinicians are interested in using AI/ML tools.
From front-line workers to back-office staff, every stakeholder in the healthcare continuum can leverage AI technology to enhance healthcare quality. In fact, ML models are already being used to collect and analyze patient data at scale, thereby allowing healthcare organizations to reduce human error and streamline physician workflows. The significant benefit of this technology is that clinicians have more time for hands-on patient engagement.
AI is also able to ingest patient health information to fuel ML models that can identify correlations between assumed diseases and other patients’ symptoms. This gives doctors a stronger understanding of underlying disease patterns and can ultimately lead to better health outcomes for patients.
However, the true power of AI in healthcare will only be realized if ML models function properly and as predicted over the long run. Ensuring consistent and accurate model performance requires constant monitoring–something that healthcare organizations need to prioritize.
Why is Model Monitoring Important?
An ML model is a type of algorithm that combs through large volumes of data to find patterns or make predictions. As real-world inputs change that data (for example, when the pandemic first started and made prior patient data obsolete), the models inevitably decay. Such changes can render ML models inaccurate, useless, and potentially unsafe—particularly for medical practitioners and doctors.
Without ML model monitoring, healthcare practitioners can end up blindly following faulty predictions or patterns that don’t exist. Model monitoring is about identifying changes and asking, “Is something different in the world that might be invalidating our original assumptions that were used to train the model?” Through monitoring, teams can detect distributional shifts, for example, in the way that doctors describe similar symptoms. With this deeper insight, healthcare practitioners can improve accuracy related to recognizing patterns, as illness variants change and spread.
ML model monitoring is especially important for applications involving natural language processing (NLP) and computer vision (CV)–both of which work with forms of unstructured data. The former refers to the ability of AI to understand human language as it is spoken and written. The latter culls meaning from digital images, videos, and other visual inputs.
A majority of healthcare data is stored in physicians’ notes, X-ray images, CT scans, and similar forms. Therefore, being able to mine large volumes of unstructured data is immensely valuable for modern healthcare organizations, which means they’re often deploying AI tools with NLP and CV capabilities. Monitoring with NLP and CV gives healthcare providers the ability to identify changes in data that occur in doctors’ notes and other unstructured forms (e.g., text, images, videos) where it was previously hard to measure change.
Let’s take COVID-19 as an example of the role model monitoring with NLP and CV can play in healthcare. Say a developer is creating a model that ingests doctor’s notes and uses the data to identify patterns of disease for more precise illness detection. Prior to COVID-19, if a doctor’s notes involved a patient who had flu-like symptoms, the model would identify this as being a case of the flu. However, after the pandemic hit, models had to be retrained with the knowledge that flu-like symptoms could also be indicative of a case of COVID-19. Without this monitoring and retraining, the model wouldn’t reflect the current state of the world or enable better decision-making.
As a result, many are turning to solutions like Model Performance Management (MPM) that can monitor every model in pre- and post-production to help practitioners make more efficient, less error-prone health decisions. The most sophisticated MPM platforms help engineers monitor NLP and CV and pick up on nuanced data drift quickly.
Healthcare has come a long way in operationalizing AI technologies to enhance physician efficiency and improve patient outcomes. However, in order to successfully leverage ML, healthcare organizations will need to monitor their models to ensure they’re functioning correctly. With the right tools in place that can automatically identify and alert users of changes in both structured and unstructured data, healthcare practitioners will be able to confidently deploy models that could change the world of medicine for the better.
About Krishnaram Kenthapadi
Krishnaram Kenthapadi is the Chief Scientist of Fiddler AI, an enterprise startup building a responsible AI and ML monitoring platform. Previously, he was a Principal Scientist at Amazon AWS AI, where he led the fairness, explainability, privacy, and model understanding initiatives in Amazon AI platform. Prior to joining Amazon, he led similar efforts at the LinkedIn AI team, and served as LinkedIn’s representative in Microsoft’s AI and Ethics in Engineering and Research (AETHER) Advisory Board.