Health and Explainable AI Podcast

In this episode of the University of Pittsburgh’s Health and Explainable AI Podcast, host Jordan Gass-Pooré speaks with Francesco Lescai, PhD, Associate Professor of Bioinformatics and Head of the Computational Genomics Laboratory at the University of Pavia. Reflecting on his career transition from wet-lab immunology to computational genomics, Francesco discusses his work, explainable AI and the necessity of interdisciplinary training across computer science, genetics, and health informatics to solve complex human disease traits.

The conversation focuses on his lab's SIEVE framework (Sparse Interpretable Exome Variant Explainer), a deep learning framework that advances explainable AI by shifting the paradigm from post-hoc model auditing to using inherently interpretable architectures as active engines for biological discovery. For students and practitioners in explainable AI, his work offers practical insights into model development, demonstrating how training "null" baseline models ensures attributions reflect true biological signals over computational noise, how annotation-ablation protocols measure a model's reliance on raw genomic architecture versus pre-existing labels, and how capping GPU memory limits power consumption while actively measuring CO2 footprints to keep XAI sustainable. Francesco also highlights the necessity of making research openly available for external critique across disciplines to catch architectural blind spots and drive healthier, more collaborative science.

Aligning with the mission of the University of Pittsburgh's Health and Explainable AI (HexAI) Research Laboratory, Francesco offers insights for students and computational researchers on navigating sparse clinical datasets, developing robust machine learning architectures for rare diseases, and advancing inter-institutional sample sharing networks. Reflecting core priorities of Pitt's Computational Pathology and AI Center of Excellence (CPACE), the discussion illustrates how explainable machine learning and sensor-driven molecular profiling advance the frontiers of computational pathology and precision diagnostics. He closes the episode by encouraging students to build transparent, interpretable models for genomic medicine

Host: Jordan Gass-Pooré | Producer: Brent Phillips | Guest: Francesco Lescai, PhD (Associate Professor of Bioinformatics & Head of Computational Genomics Laboratory, University of Pavia)
Topics: Computational Genomics | Explainable AI (XAI) | SIEVE Framework | Variant Interactions | Data Governance & Privacy | Sustainable Computing (CO2 Footprint)
Institutions: University of Pittsburgh | University of Pavia | Norwegian University of Science and Technology (NTNU)
Links:
[YouTube] https://www.youtube.com/watch?v=MxQ3LgrPxZs&t=27s
[Pitt HexAI Lab] https://pitthexai.github.io/
[SIEVE Project] https://lescailab.github.io/sieve-project/

In this episode of the University of Pittsburgh’s Health and Explainable AI (HexAI) Podcast, host Jordan Gass-Pooré speaks with James Grana, Vice President of Value-Based Care Programs, Innovations, and Contracting with BlueCross BlueShield of South Carolina, and his colleague Andrew Fogner, Epidemiologist and Manager of Data Science and Strategic Analytics. Exploring a new vantage point for the podcast, the conversation examines the integration of artificial intelligence and advanced analytics into healthcare from the perspective of health insurance providers. James and Andrew reflect on the unique position that payers occupy within the healthcare ecosystem, highlighting how health insurance providers leverage longitudinal, population-level data across multiple health systems and care settings to evaluate total cost of care, patient outcomes, and health equity.

Bridging operational data science with clinical decision support, the discussion outlines the distinct roles of predictive models in risk stratification and generative AI in accelerating administrative workflows, documentation and care management outreach. The guests address the growing "arms race" between payers and providers regarding AI adoption for coding, approvals and contract review, while emphasizing the imperative of maintaining "humans in the loop" to mitigate hallucination risks and ensure trust. Crucially, the dialogue connects explainable AI directly to value-based care, illustrating how transparent and interpretable models build clinician confidence, support proactive population health interventions and ensure AI tools drive meaningful health improvements rather than fee-for-service volume.

Aligning with the mission of Pitt’s HexAI Research Laboratory, James and Andrew offer insights for students and computational researchers on addressing sparse data, establishing AI governance frameworks and evaluating real-world model impact. Reflecting core priorities of Pitt's Computational Pathology and AI Center of Excellence (CPACE), the discussion illustrates how explainable machine learning and sensor-driven molecular profiling advance the frontiers of computational pathology and precision diagnostics. They close the episode by inviting academic research labs to partner with payers in establishing independent model evaluations, acting as a trusted "Consumer Reports for AI” and developing an ethical moral compass to prevent fraud, waste, and abuse in AI-powered healthcare.

Host: Jordan Gass-Pooré | Producer: Brent Phillips | Guests: James Grana (VP of Value-Based Care Programs, Innovations & Contracting, Blue Cross Blue Shield) & Andrew Fogner (Epidemiologist & Manager of Data Science & Strategic Analytics, Blue Cross Blue Shield)
Topics: Health Insurance Analytics | Value-Based Care | Explainable AI (XAI) | Population Health Management | AI Governance & Ethics | Predictive vs. Generative AI
Institutions: University of Pittsburgh | Blue Cross Blue Shield of South Carolina
Links:
[YouTube] https://www.youtube.com/watch?v=etPqZQw977E
[Pitt HexAI Lab] https://hexai.pitt.edu

In this episode of the University of Pittsburgh’s Health and Explainable AI (HexAI) Podcast, host Jordan Gass-Pooré speaks with Daniel Heller, PhD, Head of the Cancer Nanomedicine Laboratory at Memorial Sloan Kettering Cancer Center and Professor at Weill Cornell Medicine. Anchoring the discussion around the convergence of computational medicine, nanotechnologies, and explainable machine learning, Daniel shares his journey from early interest in carbon nanotubes at Rice University to leading cutting-edge research in cancer diagnostics, and then goes on to provide listeners with an overview on the scope of biosensor technology and variety of internal and external biosensors being studied and applied today, ranging from micro-scale pill cameras and implantable devices down to optical nanosensors operating at one hundred-thousandth the width of a human hair.

Bridging the gap between material chemistry and artificial intelligence, Daniel details how his lab utilizes arrays of surface-modified carbon nanotubes to capture biological data. Drawing an analogy to the human olfactory system, he explains how these sensors forgo single-biomarker recognition in favor of creating multi-signal optical "fingerprints" from blood plasma, which are then evaluated by supervised machine learning models to detect early-stage ovarian cancer and differentiate brain tumor subtypes for example. Crucially, the conversation highlights how explainable AI techniques, such as feature importance analyses, SHAP, and LIME, are inverted to drive scientific discovery. By identifying which nanosensors contribute most to the model’s diagnostic decisions, researchers can wash off bound proteins to isolate rare, disease-specific biomarkers and map out the broader systemic "tumor ecosystem".

Aligning with the mission of the University of Pittsburgh's HexAI Research Laboratory, Daniel offers insights for students and computational researchers on navigating sparse clinical datasets, developing robust machine learning architectures for rare diseases, and advancing interinstitutional sample-sharing networks. Reflecting core priorities of Pitt's Computational Pathology and AI Center of Excellence (CPACE), the discussion illustrates how explainable machine learning and sensor-driven molecular profiling advance the frontiers of computational pathology and precision diagnostics. He closes the episode with a call to action for students with interdisciplinary backgrounds across programming, chemistry, and clinical biology to explore careers at the intersection of nanotechnology and explainable AI.

Host: Jordan Gass-Pooré | Producer: Brent Phillips | Guest: Daniel Heller, PhD (Head of Cancer Nanomedicine, Memorial Sloan Kettering Cancer Center & Weill Cornell Medicine)
Topics: Cancer Nanomedicine | Explainable AI (XAI) | Nanosensors | Feature Importance Analysis | Biomarker Discovery | Micro-Robotics
Institutions: University of Pittsburgh | Memorial Sloan Kettering Cancer Center | Weill Cornell Medicine
Links:
[YouTube] https://www.youtube.com/watch?v=idE_5qxqvb4&t=5s
[Guest Profile] https://www.mskcc.org/research/ski/labs/daniel-heller
[Pitt HexAI Lab] https://pitthexai.github.io/

In this special episode, HexAI producer Brent Phillips steps behind the microphone to speak with Sean McGunigal, Epic Software’s Software Development Lead for AI Research and Development. Drawing from his own daily experience navigating healthcare information management systems at a hospital, Brent guides a technical conversation on the immense role that enterprise software architectures like Epic play in orchestrating, optimizing, and embedding artificial intelligence directly into clinical workflows across global health systems.

Bridging the gap between frontier software engineering and clinical realities, Sean and Brent discuss Epic's AI development pipeline while anchoring the conversation around Epic’s technical vision where utility, explainability, and trust are paramount. Sean details how Epic’s proprietary ambient AI framework, Chart with Art, seamlessly ingests and parses acoustic conversational data to generate structured documentation, effectively alleviating cognitive burden and clinical friction so medical professionals can prioritize hands-on patient care. Sean emphasizes that to achieve user adoption, Epic's generative summaries are rigidly tied to programmatic data lineage and granular citations, forcing black-box models to prove their clinical outputs.

Aligning with the mission of the University of Pittsburgh’s Health and Explainable AI Research Laboratory and Pitt’s Computational Pathology and AI Center of Excellence (CPACE), the episode highlights Epic’s commitment to helping advance and democratize ethical and responsible AI engineering through open-source frameworks and massive data initiatives. Sean points to Nebula, Epic’s cloud-native platform deployed in Microsoft Azure, which provides a high-throughput, horizontally scalable infrastructure allowing health networks to securely deploy LLMs. Furthermore, he underscores Epic's breakthroughs with Cosmos, their multi-institution de-identified dataset, and Curiosity, Epic's custom-trained Medical Language Model designed to predict patient clinical events using autoregressive, next-token intelligence. By exploring the orchestration of complex agentic AI workflows, runtime evaluation harnesses, and cross-platform communication protocols, Sean closes the episode with a call to action for researchers to get involved and work more closely around safely advancing the frontiers of computational medicine.

Dr. Bradley Erickson, Director of the Mayo AI Lab, speaks with HexAI podcast host, Jordan Gass-Pooré in advance of the University of Pittsburgh’s annual AI Summer School program in Medical Imaging Informatics organized by Pitt's Health and Explainable AI Research Lab (HexAI) and the Computational Pathology and AI center of Excellence (CPACE). The episode simulates two different professional vantage point scenarios to help students visualize the vast, multi-dimensional landscape of artificial intelligence in healthcare and radiology.

The first half of the episode drops students directly into the vantage point of an AI expert attending a technical conference, where medical imaging informatics are being contrasted with everyday computer vision. Dr. Erickson explains how medical data often extends into multiple dimensions by incorporating complex spatial matrices and tissue properties like T1 and T2 tracking on MRIs, far surpassing standard 2D photographic pixels. He highlights why generic consumer AI tools like simple heat maps or saliency maps fall short of establishing clinical trust; while they can successfully point to where a brain tumor is, they completely fail to explain what that tumor is or why it is changing texture. Furthermore, Dr. Erickson discusses the profound challenge of "ground truth" uncertainty in medicine, explaining that training predictive algorithms is incredibly difficult because definitive biological labels are frequently masked by biological reactions or a lack of definitive longitudinal data.

The second half of the podcast episode places students into the role and vantage point of a hospital administrator, exposing students to the active economic and structural deliberations currently playing out in modern hospital boardrooms. Dr. Erickson underscores the considerations and financial constraints that hospitals contend with and explains that while new narrowly focused diagnostic AI tools are attractive, the most immediate return on investment for hospitals often comes from practical, language-based text summarization and ambient patient recording systems. Crucially, this administrative perspective teaches students that the health industry desperately needs supportive roles beyond traditional doctors and researchers, such as AI project managers, integration specialists, and governance officers who can oversee model confidence and decide exactly when to adapt AI solutions or pull failing applications or algorithms back.

Dr. Erickson emphasizes that entering this revolutionary field requires a willingness to learn through iteration, push back on assumptions, and manage the critical intersections of technology, safety, and human care. Through an open exploration of technical hurdles and administrative realities, the episode provides a rich conceptual primer for AI Summer School participants designed to cultivate critical thinking informing views on AI in medical imaging, hands-on project development and coding.