From ICU Nurse to Healthcare Tech CEO: Angela Adams RN on Solving Missed Follow-Ups with AI

In the complex ecosystem of modern healthcare, the most dangerous point in a patient’s journey is often the transition between diagnosis and action. For Angela Adams, RN, BSN, CEO of Inflo Health it is a matter of life and death that drove her from the bedside of a Cardiothoracic ICU to the helm of a technology company. Adams, a former nurse at Duke University Medical Center and a veteran of the medical device and startup worlds, sat down with host Cole Lyons to discuss a pervasive but solvable crisis: the “incidental findings” that slip through the cracks of our fragmented healthcare system.

Her journey highlights a critical pivot in health tech moving away from tools that simply generate more data and toward solutions that automate the actual work of care coordination. This discussion explores how specialized AI, high-reliability principles, and a “human at the top of the pyramid” philosophy are finally closing the loop on patient safety.

The Silent Crisis: Why Do Critical Findings Slip Through the Cracks?

The primary reason patients fall through the cracks is that health systems excel at acute, immediate treatment but often lack the infrastructure to track incidental findings across different care episodes.

This disconnect is best illustrated by the tragedy that inspired Inflo Health’s founding. Adams recounted the story of a colleague, Jill, who visited the ER for appendicitis. While the hospital successfully treated her acute surgical need, a CT scan revealed a suspicious breast lesion—an incidental finding unrelated to her appendix. Because there was no automated safety net to catch this detail, the finding was noted in a radiology report but never acted upon. Ten months later, a routine mammogram rediscovered the mass, but it was too late; the cancer had metastasized to her brain.

This is not an isolated anecdote, It is a systemic failure that Adams and her team validated with hard data:

• 50% of radiology follow-ups documented in reports are missed completely.
• 80% of follow-up volume (incidental findings outside of regulated breast and lung screening programs) is largely ignored by existing tools.
• Reliance on “Heroic Effort”: Systems currently depend on clinicians manually bridging gaps between departments, a method that is statistically destined to fail in a high-volume environment.

“It would’ve taken a heroic effort on behalf of the clinical teams, which is a lot of times what we expect out of our clinicians. We expect them to make a heroic effort to get work done every single day. And in this particular case, she just completely fell through the cracks.”

Why Traditional EHR “Worklists” Are Not the Solution

Current Electronic Health Records (EHRs) fail to solve this problem because they primarily generate static worklists that increase administrative burden rather than automating the closure of care loops.

For years, the industry assumed that if AI could detect a problem and put it on a list, the problem was solved. However, Adams argues that this approach fundamentally misunderstands clinical reality. A worklist does not schedule a patient. A worklist does not educate a frightened patient about what a “nodule” means. A worklist merely transfers the burden of action onto an already overwhelmed administrative staff who are forced to “dial for dollars” to track down patients.

Adams emphasizes that true reliability requires moving beyond identification and into orchestration.

• Identification: AI finds the keyword in the report.
• Orchestration: Automation routes the finding to the correct provider, initiates the order, tracks the scheduling, and educates the patient via text.

By focusing only on detection, many vendors are simply highlighting problems without offering the manpower to fix them.

“It irritates me… All we do by turning this on is increase our own liability. Increase the patient safety risk, increase the staff burden… Whereas a lot of companies are out there turning all of this on, and the data is just sitting there in piles with nobody to work on it.”

The “Human at the Top of the Pyramid” Philosophy

Inflo Health changes the clinician’s role by automating routine administrative tasks, allowing providers to focus solely on complex cases that require human empathy and judgment.

One of the most compelling aspects of Adams’ strategy is her rejection of the idea that AI replaces clinicians. Instead, she advocates for a structure where automation handles the 80% to 90% of straightforward cases—scheduling, standard notifications, and tracking. This promotes the human clinician to the “top of the pyramid.”

In this model, nurses and care navigators are no longer data entry clerks transferring information between spreadsheets. They are elevated to perform the work they were trained for: navigating complex patient emotions and managing difficult diagnoses.

The Impact on Clinical Staff:

1. Reduction in Burnout: Eliminating the “fishing” for information.
2. Return to Purpose: spending time on direct patient interaction rather than logistics.
3. Efficiency: Handling higher patient volumes without proportional staff increases.

“You’re promoting the human to the top of the pyramid. So it’s like AI at the bottom, automation second, and now we’re using our clinicians and their knowledge and their wisdom… as the orchestrators of the AI and the automation so that it brings joy back into their day.”

Why Custom “Mini-Models” Beat Generic LLMs

Building a custom Small Language Model (SLM) is necessary for radiology because generic Large Language Models (LLMs) are cost-prohibitive at scale and prone to hallucinations regarding specific clinical nuances.

In the current tech hype cycle, the temptation to use off-the-shelf LLMs (like GPT-4) for every problem is high. However, Adams details why Inflo chose the harder path of building proprietary, domain-specific models.

• Cost and Consumption: Health systems generate millions of imaging reports. Pinging a commercial API for every report is fiscally unsustainable and computationally heavy.
• Accuracy and Hallucination: A generic model might identify a follow-up but struggle with the “logic tree” of clinical decision-making (e.g., identifying a nodule, checking the size, checking smoking history, and determining the specific guideline-based action).
• Speed: “Mini-models” trained specifically on radiology ontologies are faster and easier to deploy within a hospital’s secure environment.

This decision highlights a maturing of the AI market: the move away from “generalist” AI toward highly specialized, vertical-specific intelligence.

“If I build all of these like mini models and I build our own language structure, it’s not gonna be considered a large language model, but it will be considered a radiology specifically language model.”

Beyond SaaS: The Necessity of “People, Process, Then Tech”

Implementing high-reliability change requires a partnership model that addresses people and processes before layering on technology, ensuring the underlying workflow is sound.

Perhaps the most critical lesson for healthcare leaders in this episode is the admission that technology alone cannot fix a broken system. Adams describes their partnership with the American College of Radiology (ACR) “Empower” program, which utilizes A3 problem-solving methodologies to identify failure points before software is even installed.

If a hospital does not have enough CT scanners to handle the influx of patients that the AI identifies, the software will fail. Inflo Health’s approach involves consulting on the operational roadmap—predicting volume surges and helping hospitals staff accordingly.

The Implementation Pillars:

• People: Who is responsible for the finding?
• Process: How does the data move from Radiology to the ordering provider?
• Technology: The automation layer that enforces the process.

“Technology alone is never going to be able to fix everything in a health system. It has to be like people, process. Then it’s all layered on tech that can do the automation.”

The Future: Agentic AI and Empathetic Automation

The next frontier involves “Agentic AI” capable of holding nuanced, empathetic voice conversations with patients to navigate complex diagnoses like cancer. Looking toward late 2024 and beyond, Adams predicts the rise of Agentic AI—autonomous agents that can handle bi-directional voice and text communication. However, the challenge in healthcare is vastly different from retail or customer service.

An agent calling a patient about a missed prescription is simple. An agent calling a patient about a suspicious lung nodule requires a high degree of “emotional intelligence” programmed into the bot. The system must be able to answer questions like “Do I have cancer?” or “What is a PET scan?” without causing panic or violating medical ethics.This evolution from text-based automation to voice-based agents represents the next leap in efficiency, allowing health systems to manage population health at scale without hiring armies of call center staff.

“It’s not that it’s easy… your bot has to be able to have really sensitive conversations about a PET scan that has maybe the opportunity that the patient has cancer, like teaching a bot to have that level of conversation.”

Actionable Insight

For healthcare executives and digital transformation leaders, Angela Adams’ journey underscores a pivotal shift in strategy: Stop buying detection; start buying closure.

When evaluating AI tools, look beyond the accuracy of the algorithm (the “detection”) and interrogate the workflow (the “action”). Does the tool simply add rows to a spreadsheet, or does it autonomously advance the patient to the next step of care? Audit your current incidental finding workflows. If your process relies on a human to manually transfer data from a report to a registry, your system is vulnerable. The future belongs to platforms that treat the “finding” not as the end result, but as the trigger for an automated, high-reliability workflow.