The health field is increasingly attracting AI players, including Microsoft.
Last week, the company announced expansions to Fabric, the data analytics platform it unveiled in May, to allow Fabric to perform more analytics on multiple types of health data. Microsoft also announced new services for its Azure Cloud Computing service to, among other things, use large language models as medical assistants.
“We want to build this unified, multimodal database in Fabric One Lake, where you can unify all these different data modalities so you can reason about that data, run AI models, etc.,” said Umesh Rustogi, general manager of Microsoft Cloud for Healthcare, in an interview with ZDNET.
“The creation of patient cohorts based on criteria drawn from imaging results and clinical results”
The multi-modality trend, which ZDNET explored in a feature article on AI this month, is increasingly important in healthcare, Rustogi said. “Many clients believe that if you combine multiple data modalities, you can gain new insights that are not possible by researching a single data modality,” Rustogi said.
Umesh Rustogi, general manager of Microsoft Cloud for Healthcare. Microsoft
Examples of multi-modalities include “creating cohorts of patients based on criteria drawn from their imaging results and their clinical results. And this is not very easy to achieve today” , did he declare. Mr. Rustogi cited as an example a study published in 2020 in the prestigious journal Nature. This article provides an overview of “data fusion” techniques that can be “applied to combine medical imaging and electronic health records (EHRs).”
Another of Fabric’s new capabilities is a “de-identification service,” which uses machine learning to clean clinical data to mask patient identities in data such as doctors’ notes. “The industry has had a very difficult time solving the question of how to take these unstructured clinical notes and de-identify them in a way that keeps them useful to the research community,” Rustogi said.
What sources should we use for AI in the health field?
Mr. Rustogi’s colleague Hadas Bitran, head of Microsoft’s Health AI and Health and Life Sciences division, discussed several new artificial intelligence offerings offered by Azure web services.
The Azure AI Health Insights offering consists of pre-built machine learning models. Three models are offered in preview:
- Patient timelinewhich “uses generative AI to extract key events from unstructured data, such as medications, diagnoses and procedures, and organizes them chronologically to give clinicians a more accurate view of a patient’s medical history in order to better inform care plans”;
- Simplifying clinical reportswhich “uses generative AI to give clinicians the ability to take medical jargon and convert it into plain language while preserving the essence of clinical information so it can be shared with others people, including patients”;
-
Radiology insightswhich “provides quality controls through feedback on errors and inconsistencies. The model also identifies follow-up recommendations and clinical findings in the clinical documentation with measurements (sizes) documented by the radiologist.”
These three models are in addition to several pre-built models already proposed for clinical trials and for phenotype-based models in oncology.
A new offering called Azure AI Health Bot uses large language model (LLM) technology to extract answers and answer questions from sources such as a healthcare organization’s own database. “The idea is that this service helps customers create specialized co-driving experiences,” Mr. Bitran told ZDNET in the same interview with Mr. Rustogi.
“What is also interesting is that it is possible to create a cascading effect,” added Bitran. “So you can use your own sources and, if they don’t give anything, you can also provide answers based on credible sources, and then, if the credible sources don’t give anything, you can just fall back on a generic answer.”
Of course, there is currently a lot of skepticism about the use of generative forms of AI, such as large language models, in sensitive practices such as healthcare. What does Microsoft think about these concerns?
Hadas Bitran, head of Microsoft’s Health AI and Health and Life Sciences division. Microsoft
“Great language models need to be complemented by something else to give good results”
“Large language models need to be complemented by something else to give good results,” said Hadas Bitran. “For every model that we create, if we use large language models, they will always come with safeguards specific to the health sector,” added Bitran.
“One of the most interesting approaches is to use smaller models and rule-based models, in a hybrid model with LLM,” added Bitran.
For example, in the prebuilt model for clinical reporting simplification, “we don’t just ask the language model to explain. We also implement a lot of preprocessing and postprocessing logic that allows us to take the result of the simplification, measure it against the performance metrics of the simplification,” explained Mr. Bitran. “We then cross-reference to see if the results are actually a simplification of the source, or if there are fabricated or missing elements.”
“Our models are not intended to replace the doctor”
“This responsible AI framework is not just about privacy, security, accessibility and transparency, etc. “It’s also about accuracy, accountability and fairness.”
“Finally, our models are not intended to replace the doctor,” added Mr. Bitran. “They aim to provide clinicians with tools that lighten their workload and help them in their work.”
Source: “ZDNet.com”