Keeping every patient healthier as fast as care demands

Based in Chicago, Illinois, Duly Health and Care is the largest multi-specialty, physician-owned practice in the United States. Duly operates two business models: a fee-for-service, where they bill per visit, and value-based care, in which Medicare provides a fixed payment per patient based on their health profile. As the organization scaled, three critical challenges threatened the company’s financial performance and the quality of care: providers were referring patients outside Duly's network, readmission rates were dangerously high, and business teams had no easy way to access or act on their own data. Duly took proactive steps to resolve these obstacles, invested in the Databricks Data + AI Platform and reduced readmission rates by 29%.

How did Duly want to use data and AI to improve patient care?

Duly Health and Care is a physician-owned organization where providers who meet certain criteria can become shareholders, making the company's success personally meaningful to the people delivering care. Duly operates across two business models. The first is traditional fee-for-service: a patient visits a clinic, a physician provides care, and the physician bills insurance for each service rendered. The second is value-based care, a U.S. Medicare Advantage program designed to increase patient health while decreasing costs across the healthcare system. Under VBC, Medicare calculates a risk score for each patient based on their diagnoses, age, social determinants of health, and location, then provides Duly a fixed monthly stipend to care for that patient. Duly effectively becomes the payer. If a patient ends up in the hospital, the company absorbs those costs.

To keep patient care within the Duly network, the team wanted to create a "Physician Finder" that lets providers quickly search and refer patients to the right in-network physician. Providers could search by specialty, procedure, language, gender, tenure and more. Using a conversational AI interface, they’d type or speak natural-language queries, such as "find me a gastroenterologist taking new patients in Elgin," and receive answers in plain English on mobile or desktop.

Duly also planned to build a readmission prediction model to thoroughly analyze whether a patient would be readmitted to the hospital within 30 days of discharge. "Under value-based care, we lose visibility into a patient’s care if they see someone outside of the Duly network. With every preventable readmission, we're the payer, so that hits our bottom line directly. We wanted to explore AI approaches that were all about keeping patients healthier by keeping them within our network and care,” said Jim Ritchhart, Executive Director, Data and Analytics at Duly Health and Care.

Duly planned to expand their AI strategy more broadly and bring self-service to data spaces. In these designated spaces, any team member could ask questions of live data in plain English. Starting with their flagship space, Patient 360, Duly covered 200-300 columns of VBC patient data — readmission history, appointment dates, medical costs and condition flags. Next, the Provider Space tracked key productivity and billing metrics, while the Finance Space handled claims, denials and invoices. Lastly, the HR Space pulled headcount and FTE data directly from Workday, and the Appointments Space gave the scheduling team visibility into patient appointment history and upcoming provider schedules.

Why were Duly’s referrals leaking and readmissions rising?

For a physician-owned organization, the stakes around data are high. Since providers regularly referred patients outside Duly's network the costs added up. The tool meant to solve this — the initial “Physician Finder” — was built on an Excel sheet that didn't update in real time, with a static dropdown interface that gave providers little reason to trust or use it. On the clinical side, Medicare actively judges organizations like Duly on readmission rates, and with Duly's sitting around 14%, there was no way to proactively identify which patients were at the highest risk before they ended up back in the hospital.

The lack of data access across the business compounded these issues. Non-technical teams had no self-serve way to get at their own data — the workflow was entirely manual, requiring exports, Excel filters and tickets to the data team for anything beyond what was already built. Underpinning it all was an infrastructure problem: Duly's legacy SQL Server warehouse pulled roughly 90% of its data from Epic, required handwritten stored procedures that delayed new data requests by up to two weeks and ran full daily refreshes on records that never changed. Data was siloed with no unifying layer, and some of it sat on servers from 2007 that hadn't been touched in nearly two decades. After spending a day at Microsoft evaluating options, Databricks emerged as the clear path forward, since it was flexible enough to unify Duly's fragmented sources, fast enough to replace handwritten stored procedures and scalable enough to grow with the organization.

What did it take to replace a 17-year-old infrastructure?

Duly began their migration to Databricks with Delta Lake. The healthcare company unified nearly 60 data sources including Epic, D365, Workday, call center data and historical EMRs into a single, reliable layer. Built on Medallion architecture, data flowed through three progressive stages: bronze (raw ingested data), silver (cleaned and validated data) and gold (business-ready data). Using Delta Lake, Duly enabled incremental data loads rather than full daily refreshes for patient records that rarely (or never) changed. After the data foundation was established, Unity Catalog was layered on top of Delta Lake. For an organization of Duly's complexity, knowing where data came from, who could access it and what it meant was critical. Unity Catalog handled all of it: tracking column and table definitions, controlling group-based permissions, organizing table and query lineage and managing vendor access to isolated catalogs to keep sensitive data protected.

By providing access to clean, governed data, Databricks Workflows and Jobs brought the same order to code execution, giving the team a centralized place to schedule, orchestrate and monitor pipelines across the organization. The “where did this come from?” conversations that used to slow everything down disappeared almost entirely. Additionally, the technical teams at Duly benefited from Databricks SQL, an SQL-based interface that allowed them to query data directly. Powering it was serverless compute, which eliminated the need to keep clusters running constantly during business hours. Previously, Duly had to pre-warm infrastructure just so teams wouldn't face delays when they needed data — serverless changed that, spinning up warehouse resources on demand and scaling back down when finished. The same on-demand speed extended to dashboards, where Databricks' AI/BI capabilities let the team build and deploy data quality and tracking views just as quickly.

The data science team then used MLflow to build the readmission prediction model and rapidly test and compare algorithms, such as Random Forest, LightGBM and XGBoost, across a range of hyperparameters. AutoML handled model construction inside the platform, while Notebooks made real-time tweaks. Where MLflow brought intelligence to the data science team, Genie Agents brought it to everyone else — giving non-technical users across the organization the ability to ask questions of live data in plain English and get answers back instantly. To cover the full breadth of the organization, Duly built five in total, each scoped to a specific team and dataset: Patient, Provider, Finance, HR, and Appointments. “Because Genie draws directly from Unity Catalog’s metadata, the column definitions, table descriptions and business logic were already documented in the platform, meaning training each space was significantly easier each time,” explained Jim. Through the Genie API, Duly built a custom Streamlit app for scheduled email delivery of query results, in-app graphing and downloadable exports. The result was a single platform where any team member including clinical, financial or operational teams could access, query and act on their own data without waiting on anyone else.

What did cutting readmissions and doubling productivity look like?

Duly's investment in Databricks spanned clinical outcomes, financial performance and organizational productivity, and the results included:

• 29% drop in readmission rates — from 14% to 10% — representing $4.5M in annual hospital charge savings
• 30% reduction in data platform costs compared to SQL Server on a like-for-like basis
• 3-4x more data stored and ingested faster across nearly 60 unified sources
• 2x project output in 2025, with ~106 projects completed with the same headcount that delivered 51 in 2024

Built on 100,000 historical Duly patients and achieving a ROC AUC of 89.2%, the readmission model sat at the center of the clinical impact — ranking discharge risk with enough reliability to drive daily operational decisions and match paramedics and physicians to the patients who needed them most. Because the model is trained exclusively on Duly's own patient data rather than purchased or external datasets, its accuracy exceeds anything available off the shelf.

The Physician Finder, now integrated directly into the Epic workspace and accessible from any provider's phone, has similarly closed the referral leakage gap that was costing Duly both revenue and continuity of care. The impact of Duly AI was felt beyond the organization entirely. For example, a demo of the Patient 360 Genie agent generated enough internal excitement that Duly's venture capital partners requested follow-up demonstrations for their own business intelligence unit and a second portfolio company. "We decommissioned servers dating back to 2007 and replaced costly third-party storage contracts by bringing that data in-house. For the first time, non-technical teams across VBC, finance, HR and operations could query live data without submitting a single ticket,” concluded Jim.