Electronic health record (EHR) adoption was driven, in part, by a recognition of the power of data. Given the complexities of healthcare informatics, though, the full potential of health data has been largely stifled—until now. Thanks to recent advancements in artificial intelligence (AI) and machine learning (ML), we can finally turn clinical information into a truly powerful asset that lessens technology burdens and opens up capacity for more human interactions.
At Altera Digital Health, our developers have been experimenting with AI and ML applications to push the limits of data use in healthcare. Here’s a look at some of those already in place or in development from our CareInMotion team:
Surfacing data in seconds
Traditional keyword search for clinical documents has been largely ineffective as users must know precisely which terms to query—a difficult task when facing a mountain of notes, unstructured documents, progress reports and discharge summaries. Now, with generative AI, we are enabling more streamlined search with contextual meaning. Instead of manually reading a lengthy discharge summary to find necessary equipment, a user can simply ask the system to create a list of discharge orders. In addition to saving clinicians time, this also helps payers ensure members/beneficiaries receive necessary equipment, reducing the risk of readmission, poorer outcomes and higher costs.
Additionally, we’ve applied a chatbot to an interface of the dbMotion™ Solution that is fed a normalized version of the entire patient record. Instead of clicking through windows, the clinician can simply ask the chatbot a question about the data and save prompts for future use. While all clinicians may find this useful, you can really appreciate its effectiveness in a setting like the emergency department, where a provider may not be familiar with the patient but needs to act quickly and decisively.
Bringing structure to unstructured data
Unstructured data has been a white whale in health IT, but now, with generative AI and natural language processing (NLP), we can automate the extraction of discrete data points from unstructured documents. Information like family history, social determinants of health (SDOH) and allergies can be pulled from progress reports, discharge summaries or notes and automatically added to the patient’s structured chart. This not only enriches the patient record, but also significantly cuts down on manual data entry and medical coding.
In collaboration with a major national health research institution, we are building a rare disease identifier because the institution deals with a lot of unstructured text. Its researchers have been forced to comb through documents for patient histories and other information. Applying AI, we can extract discrete data from those documents and map them to standard codes. Documenting rare diseases can be challenging as they may not have ICD-10 codes yet, but we can put the data in a standardized format so it’s usable for analytics or other research tools.
Identifying patterns, from patients to populations
Alongside a national health ministry client, we are developing ML algorithms to predict negative outcomes for patients with chronic diseases. By training models on de-identified data from specific patient populations, we can identify risk indicators in real-time. For example, the system could flag a hospitalized diabetic patient with rising blood pressure as being at high risk for an increased length of stay or readmission, allowing the care team to intervene proactively.
Working with a Canadian provincial health agency, we have also demonstrated the ability to identify potential public health issues, such as opioid abuse. Using a synthetic dataset, we successfully identified scenarios indicative of “opioid shopping,” such as a patient receiving opioid prescriptions from three different providers in a 60-day period. From that data, we generated heat maps to visualize areas with higher instances of potential misuse. The AI can also suggest new lines of inquiry, helping public health officials and providers frame new questions to better understand and address complex challenges.
Ongoing evaluation and improvement
Finally, we are building a platform with our research institution partner to continuously evaluate the performance of different AI models so we can determine which ones are most effective at performing specific tasks. By tracking performance, gathering feedback and consistently improving the models, our partners can trust the data and insights derived from it. This concept of an evaluation platform is something we intend to apply universally to maintain transparency and trust across our AI processes.
For every data point, there is a patient. And for every patient, there are people who care for and about them. At Altera, we care about all the people that make up the healthcare system. And we have only begun to scratch the surface of making that system more helpful and more human.
Learn how Altera’s AI-enabled solutions are taking on the burden of data, documentation and decision support here.
