AI-powered lung cancer screening
Pixel AI-powered platform that integrates with imaging devices and EMRs to enable incidental detection of lung irregularities, improved accuracy and speed of diagnosis of lung conditions (COPD, ILD, nodules, etc.), and enhanced longitudinal monitoring of identified areas of concern to inform therapeutic responses.
I was responsible for product strategy and conceptual design of this platform, predicted to save thousands of lives through enhanced early detection and treatment of lung cancer and estimated to be responsible for $300M in annual recurring revenue for the client by 2027.
My responsibilities:
Generative research
Product strategy
Conceptual design & testing
Project length:
9 weeks
280%
Increase in rate of Stage I detection at pilot sites
+59pp
Improvement in 5-year survival rate for Stage I diagnoses (vs stage IV)
$300M
Estimated ARR for product line by 2027
~$30M
Projected lifetime downstream value realized for pilot site in 1 year
Context
People with early-stage lung cancer have a reasonable shot at survival, given prompt treatment. Sadly, only 16% of patients get a timely diagnosis, meaning thousands of years of life are lost. Even worse, many of these cancers could have been detected earlier had incidental chest scans (for other medical purposes) been screened for lung nodules or other areas of concern.
How might we leverage incidental scans to accelerate identification and timely treatment of lung cancer?
How might we enable pulmonologists to more confidently monitor large volumes of patients in large-scale lung cancer screening programs?
How might we leverage AI to enhance and enrich clinicians’ diagnoses and decision-making?
Concept
In light of the above questions, I designed a system that integrates an AI engine directly into the hospital imaging workflow. It works by:
Automatically analyzing all eligible images with the AI
Presenting any AI findings to the radiologist during their routine image review
Triggering follow-up by a pulmonologist for areas of concern confirmed by the radiologist
Additionally, I created two separate but interconnected applications for radiologists, pulmonologists, and nurse navigators. These applications allow them to review AI outputs and manage patients with identified areas of concern. The applications are designed to connect with electronic medical records (EMRs) and provide workflow support to ensure a seamless experience for clinicians.
MyPatient
Pixel-AI analyzer and viewer of imaging to automatically identify lung nodules, tumours, and other lung abnormalities in order to flag potential areas of concern for follow-up on any chest scans taken and support faster, more confident diagnosis and monitoring by pulmonologists.
MyProgram
EMR and imaging platform-integrated lung cancer screening patient management platform with automated patient outreach and follow-up.
Notable highlights of both products include:
Continuous AI-driven analysis
MyPatient facilitates incidental discovery of lung cancers and other lung abnormalities by running continuously on all chest scans that meet imaging requirements. This happens automatically alongside traditional radiologist review.
Modular pixel-AI ecosystem
Core pixel-AI capabilities are modular components that can be added, removed, and tailored to meet the diagnostic thresholds approved for each institution. As new pixel-AI capabilities emerge, they can be integrated into a common workflow, reducing churn and need for retraining.
Seamless EMR integration
Leveraging additional voice-to-text and automated medical consultation note-taking capabilities offered by the client, MyPatient and MyProgram support automated tracking of findings and consultation in industry-leading EMRs, accelerating clinician workflows and reducing administrative burden.
Streamlined workflows across specialists
With designated interfaces and structured, automated hand-offs between radiologists, pulmonologists, nurse navigators, and oncologists, MyPatient and MyProgram ensure patients don’t fall through the cracks while reducing the burden of transfers of information and orchestration of activities across these roles.
Work plan & approach
Given client timelines and their desire for a “bias towards action”, I opted for faster iterations using a design sprint methodology to explore and evaluate concepts and build excitement with high-fidelity mockups before more detailed refinement of workflows.
How I delivered impact
Bias to action: High-fidelity prototyping from day 1
When I initially presented a more structured work plan to the client, they expressed concerns about the extended time allocated to research and lower-fidelity design and testing. Upon further investigation, I discovered that the product team leadership was under pressure from executives to demonstrate a compelling vision for additional funding. To address this, I used a modified Material design system and a hypothesis-driven approach to create an initial high-fidelity prototype. Additionally, I outlined specific, valuable use cases to support an executive presentation within a week. The positive response to this approach allowed me to undertake more robust research and gain access to observe tumour boards and daily activities with interventional pulmonologists affiliated with the client.
Implementation readiness: Early-stage coordination with technology, marketing, & operations
While this was conceptual work, the client was eager to translate the work into a working product quickly. To help achieve this, I collaborated with various teams, including marketing, sales, software development, AI licensing, and customer operations. My goal was to ensure that the work aligned with existing processes and technologies, making it easier for new and existing clients to adopt. I also considered any customizations and integrations commonly requested by customers. Working closely with the client product team, I evaluated EMR integration points using developer “sandbox” tools and provided implementation insights and recommendations, along with experiential acceptance criteria for preliminary user stories.
Real-world impact testing: Wizard of Oz and pilot site deployments
A major challenge for this sort of product is to ensure that specialist users can trust the algorithms’ outputs and can effectively evaluate the reliability and utility of those outputs without needing to redo the work. Since there wasn’t a functional user-facing platform yet, but we had AI capabilities, I proposed conducting “Wizard of Oz” testing. In this approach, I processed a small subset of scans provided by an affiliate pulmonologist. The resulting AI outputs were embedded into an interactive prototype. This allowed me to assess the necessary metadata for user confidence with a “real” usage test. By including real data and offering genuine insights in the prototype, I achieved valuable results. Clinicians even forgot it was a prototype and interacted with it as if it were real. This approach helped me identify workflow improvements and enhance interactive components. Simultaneously, we evaluated business outcomes and gathered clinician perspectives from several pilot field sites. These sites were using manually run AI analyses, which surfaced information to clinicians engaged in real patient treatment.
Research efficiency: Augmented understanding through robust secondary research
We faced challenges in interviewing specialist medical practitioners due to limited availability and high costs. To make the most of the limited time with these clinicians, I focused on deeper secondary and internal research. I conducted secondary research from two perspectives:
Patient Understanding: I studied patient guides and clinic descriptions from various lung clinics (independent specialists, lung cancer screening programs, hospital network pulmonology, and oncology clinics). This helped me understand how clinics explain diagnosis, monitoring, and treatment to patients, informing my understanding of how clinicians engage with patients and findings.
Clinician Perspective: I reviewed pulmonology textbooks and reports from the American Lung Association (ALA). This gave me insights into how clinicians are taught to think through diagnoses and treatment.
I also interviewed colleagues who had experience as pulmonologists, radiologists, or oncologists. This helped me understand clinicians’ day-to-day realities, acronyms, and medical jargon. By focusing on norms, idiosyncrasies, and nuances, I was able to go beyond a simple mental model of general medical practice.