Microsoft Corporation

11/17/2026 | Press release | Distributed by Public on 11/17/2026 12:16

At Apollo Hospitals, clinicians use a continuous learning AI system to improve patient safety reviews

By participating in the clinicians' workflows directly, the FDE team could identify their implicit knowledge, translate it into the agent's capabilities, and continuously refine the system based on clinical feedback.

On top of daily calls with clinicians and observing the huddles, Microsoft engineers worked closely with Apollo's tech team to match the electronic health-records data already being collected by the hospital against data that was important for the governance team, to automate how the agent processes it. During the development process, engineers used synthetic data for their evaluations and testing, and not Apollo's electronic health records.

Apollo's chief information officer, Ashokkan Somuveerappan, says the FDEs performed better than traditional software engineers because they prioritized the outcome instead of only completing their individual tasks.

"They were more glued to the outcome that they needed to achieve," he says. "They would step in to get a job done. It's a different, innovative way of working-traditionally you have separate engineers for the app, the database, the UI. Now a single person can work across the different areas."

A continuous learning loop and trend-spotter

The DTH agent isn't just a productivity tool to make reviews run more smoothly. It is also a continuous learning system that ingests what the medical experts discuss at the daily huddles, using Microsoft Work IQ for context.

"The agent is reading all of this data from Fabric and actually writing things back into Fabric as well," says Edwards. "So, it becomes not just the source of data for the patient histories, but also for these cases themselves-the types of decisions that are made and the types of questions that are being asked."

That information, like Apollo's patient data, is stored in Microsoft Fabric in accordance with the hospital's data-handling standards.

This continuous learning loop-referred to as "hill-climbing" in AI-means the agent improves with use and contains Apollo's expanding enterprise knowledge of the review process. This knowledge informs the questions the agent asks clinicians, and its suggested case assessments and recommendations. The clinicians then evaluate the agent's input and make changes to it before sending the case for review.

"The number of times those questions and recommendations are getting edited is a measurement of the accuracy of the DTH agent," says Somuveerappan. "If it's able to recommend the right set of potential questions to ask, and nobody disputes those questions, the agent is picking up knowledge and context."

As the agent learns to ask clinicians better questions of an incident, frontliners too are gaining knowledge of what information is relevant for an incident review, and for patient safety.

"At the unit level, we're ready to answer questions from the leadership because the AI has generated questions to help us prepare," says Juliet Joji Varghese, Apollo's director of nursing in Chennai. "And we're also learning from what's happened at other Apollo hospitals. So, we go beyond looking at why or how something happened, to how we can prevent it in future."

Another feature of the agent is its ability to spot patterns or trends from the data it collects. This so-called cluster analysis is still a work in progress, says Dr. Sridhar, but the results are promising. One example was cases of patients developing bloodstream infections associated with central lines, which are intravenous lines placed in major veins.

"The agent flagged a cluster of central line-related infections in the bone marrow transplant unit," she says. "So the unit took action to address that issue."

Microsoft Corporation published this content on November 17, 2026, and is solely responsible for the information contained herein. Distributed via Public Technologies (PUBT), unedited and unaltered, on November 17, 2026 at 18:16 UTC. If you believe the information included in the content is inaccurate or outdated and requires editing or removal, please contact us at [email protected]