Testing Decision Support

 Role: Design lead | Team 1 Product manager, 1 engineer, 1 data insights engineer, 1 clinical lead | Timeline:January 2024-present |  Where we are now: Preparing to launch to our first wave of beta practices

The Right test for the right patient at the right time

The Testing Decision Support (TDS) solution is a point-of-care intervention embedded in OncoEMR, Flatiron Health's flagship oncology-specific electronic medical records product. The tool analyzes patient data to alert oncology providers when a patient may be eligible for molecular profiling testing, a critical step in identifying targeted treatments that can significantly improve length and quality of life for cancer patients.

I led UX strategy, research, and design for this project from early discovery through a six-month pilot across 18 oncology practices. The pilot launched in April 2025 and resulted in a 16.28% increase in testing rates over baseline and an estimated 99 additional patients receiving testing they may not have otherwise.

Curiosity

Testing Rates are falling behind

Research

Discovery

Before beginning the design process, I wanted to talk to more of our providers. If doctors were aware that testing could lead to better treatment, why were they letting the opportunity slip through the cracks? After an analysis of previous conversations conducted with providers that use Flatiron’s products in 2018 and 2020, I built a research plan to help us fill in the gaps.  I found that while the landscape of testing changes rapidly year over year, many of the same themes persisted.

Participants | 5 internal Flatiron clinical team members, 5 doctors from practices within our network (multi-specialty)

Timeline | March-April 2022

Best practices for designing at the point of care 

Because this solution was going to be deployed in Flatiron's flagship electronic medical records product, OncoEMR, I also wanted to incorporate these insights into existing design principles. These principles coalesced as a result of consistent clinician feedback about point-of-care solutions.

Keep information brief, direct and to the point

"I don't have time to read mountains of text. Give me the important information first." 

Put information in places providers already are

"I don't want to be running around to different pages and workflows constantly. I want everything at my fingertips when I'm working." 

Minimize friction, maximize engagement 

"Some of these workflows already have too many clicks. It drives me crazy." 

Be transparent and clear with all clinical information

Creativity

To start, I explored a range of options in highly trafficked pages in OncoEMR and plotted them against expected technical lift and the amount of interruption they would impose on the provider's workflow. Ideally, we wanted a solution that was a moderate to low technical lift and interruption.

Some initial feedback from our users highlighted clear winners from the initial exploration as well as some additional directions to try: 

• Providers liked the idea of a badge in the list of upcoming visits and the header of each patient's chart, but some worried that it would get lost in the shuffle or that they might experience 'badge fatigue' given that it would be one of multiple badges a patient might have

• As we predicted, providers had no interest in a popup interrupting their workflow, but recognized that it would certainly make them have to pay attention to the information 

• Providers liked the genomics page idea but were skeptical as to whether or not they'd actually remember to visit it during their busy day 

• Highlights and messages in the list of upcoming visits and 'treatment plan' were noticeable and timely, but providers worried that they weren't quite appearing in the right place at the right time

Based on this feedback, I doubled down on the badge exploration which seemed best positioned to deliver the right information at the right time without being overly annoying. I also started to explore a solution in the  ‘diagnosis and staging workflow,’ where providers classify the cancer type and severity for their patient.

Badges

Deeper explorations of the badge concept with louder colors and the ability see more details about eligibility via a clickable dropdown

A final round of feedback with some users from other places in our practices confirmed that we were honing in on the right pattern: 

• Providers liked seeing more information about eligibility incorporated in the alert

• Providers liked the idea of being alerted when they were looking at the patient's diagnosis and stage because it catches them closer to the moment when they would be able to act on the information 

• Providers liked seeing information that was prominent but not blocking

Renewed focus

After our final round of testing, the testing decision support work was paused for about a year while the team refocused on higher priority work. In 2024, the company re-prioritized the project with a devoted scrum team. I took on the role of lead designer and used my institutional knowledge of our previous work to explore how we should proceed with more resourcing. 

Demoting the badge

Since the badge was our last most promising direction, product was deeply interested in doubling down on this pattern in OncoEMR. I had two concerns with the badge: 

• Badge fatigue:  A number of workflows were trying to make use of the badge pattern, meaning a patient may have up to 5 visible at a time. This creates 'badge fatigue' in providers, who end up ignoring these colorful indicators altogether

• Mixed signals: As different teams looked to the badge to solve disparate problems, the role of the badge was being muddled. Was it intended to indicate a status? Provide an alert? Impart additional information? I was hesitant to further stretch our design system around this new use case

Instead, I started to explore designs that leveraged the energy of the badge while not sacrificing prominence and clarity. This is how I began using the banner motif. 

Introducing the banner 

My initial exploration in this space leveraged an existing pattern in our design system, the "callout." This is often used as an embedded popup in different parts of the patient's chart for contextual alerts. In order to leverage what worked about the badge, I kept these alerts close to the patient chart header. 

Refining

Anatomy of the alert banner 

I deconstructed the alert banner's structure to determine how it could best guide providers from notification to action. By breaking down its essential elements, our team could decide which specific workflows the component should enable.

Alert icon: Something unique that immediately indicates the nature of the message

Alert text: Short, generalized and to the point, both saving space and encouraging the user to seek more information 

Call(s) to action: Gives the user options about what they can do next

• See more: Leads the user to more detailed information about the recommendation so they can make an informed decision 

• Snooze: Temporarily dismisses the alert so the provider can focus on other things until they're ready to acknowledge the information 

• Go to orders: Takes the user directly to the workflow that allows them to add a molecular profiling test 

• Close: Permanently dismisses the alert

Calls to action

I tried a few different configurations for where the links to these actions should appear on the banner. My most successful iteration separated the actions from the alert text: Because each potential action might have equal importance to the user depending on the scenario, it made the most sense to cluster them so that the provider could see all their options at once.

Exploring color (cont)

While the testing decision support alert is just that, an alert, the provider can safely ignore it without causing immediate patient harm. Not wanting to dilute the meaning of more severe alerts, I anchored to the more neutral color options. Purple emerged as the most logical choice here for two reasons: (1) OncoEMR uses blue as an accent color, and I feared the alert might get lost. (2) gray is used in other places in the EMR to indicate a disabled state. I wanted the alert to pop just enough to catch the user's eye without being an eyesore.

Bringing more information to the clinician's fingertips

One of the insights that came out of our interviews is that providers like having more information about eligibility before making their decision. The minimalist design of the alert banner doesn't allow for much detail, which is why I began to explore how we could put this information in other places that were equally accessible. 

Drawing on existing patterns in our design system, I explored ways to bring detailed information about the eligibility criteria to the provider's fingertips. Short of creating a completely new page, the options with the most promise were a popover, a modal and a side bar. 

Of these three, the side bar was the most promising to me even though it was a nascent pattern for our design system. One of the principles I wanted to make sure to bring to this workflow was minimizing friction to maximize engagement. For the purposes of surfacing more information, this means the provider must not feel like their workflow is being interrupted by the information, but rather supported. The modal felt too much like an interruption to justify. While the popover was subtle and kept the user in the context of the workflow, the pattern proved hard to scale in the case of changing or expanding information.

Designing for more than one disease 

As development continued, we recognized the need to accommodate multiple diseases or bilateral diagnoses in our design. This required restructuring the sidebar to handle multiple alerts and testing recommendations. To organize this information clearly, I developed the diagnosis 'card' concept to visually separate recommendations by disease.

With the card, all information specific to the disease is contained within the boundary.

The final solution helps identify patients eligible for molecular profiling through an integrated alert system. A banner appears immediately in the patient chart after diagnosis, offering options to dismiss, order tests, or access detailed information via the sidebar. This streamlined decision support tool helps providers order appropriate tests at the right time, serving as an efficient yet powerful point-of-care assistant.

Launch

In April 2025 we launched the TDS intervention to 18 practices across our network representing 609 patients and 723 doctors. The Pilot duration was six months.

As part of the pilot we also had a control group of 12 practices chosen because they had similar sizes and testing rates as the active group. We used this reference group as well as the baseline testing rate of the active practices over the previous six months as reference points to measure the effectiveness of our intervention.

Determining Success

Our team determined that a statistically significant increase of testing over the course of the pilot was 12%. This determination was made by studying current testing rates for NSCLC and evaluating the impact of other studies that have attempted to improve testing rates.

In order to capture accurate baselines from which to measure, we took the average baseline testing rate for the 18 pilot practices during the six months prior to activating our alert and also tracked the testing rate of a reference cohort of practices. For this reference cohort, we aimed to achieve similar populations across all three groups in regards to number of patients and mix of practice size and makeup.

Challenges

Despite having a robust strategy and several tools for capturing data, collecting the right information came with some challenges:

• Inferring conversion: While the alert had a dedicated workflow for ordering a test that could be launched from engagement with it, we consistently saw many doctors using their more familiar workflows to order tests, even after seeing the alert. While we could draw a direct line to conversion for users that saw the alert —> launched the ordering workflow —> ordered a test from the alert itself, we could not technically do this for users who saw the alert —> ordered a test through a traditional workflow.

  • Mitigation: We used two metrics to illuminate conversion: One was direct (saw the alert—> ordered a test through the alert) and the other was indirect (saw the alert —> ordered a test within 24 hours of seeing the alert through a traditional workflow

• Fullstory limitations: Due to limitations with our enterprise plan, we could not capture data from all 18 practices live on the alert. This meant that the Fullstory data was incomplete and that we would miss insights and session replays coming from certain practices.

  • Mitigation: I researched each practice and came up with a recommended sample list of practices I thought would best represent our cohort. We leaned on our data in the more comprehensive Looker dashboard for metrics and I positioned Fullstory more as a qualitative research tool than a source of truth for overall metrics. I used the segments and dashboards I set up to find relevant and interesting user playback sessions that would help us generate insights about usage.

Impact

The bottom line: Alerts at the point of care can increase testing rates

Over the course of the six-month pilot, 421 of the 609 patients that were impacted had an alert triggered on their chart indicating that they were eligible for testing. Across the 18 practices where these patients were being treated, we saw an increase in testing rates of 16.28% compared to the baseline and a 33.1% absolute increase, resulting in 99 additional patients receiving testing that may not have otherwise. This well outstripped our original goal of showing a 12% increase in testing rates.

KPI Results

Over time

The graphs below show the testing rate over the six month duration of the pilot for all 18 practices (left)* compared to a reference cohort of practices of similar sizes and with a similar baseline testing rate.

Qualitative insights

I conducted post-pilot interviews with a handful of practices to understand their experiences and go beyond the numbers. Below are some of the most notable insights from these conversations.