Nia Bowers
Guru Lakshmi Priyanka Bodagala, Pharm.D, is a Health Informatics Analyst and Digital Health Specialist based in San Francisco, California. With advanced credentials spanning a Doctor of Pharmacy from Rajiv Gandhi University and an MSc in Digital Health Informatics from the University of San Francisco, Bodagala’s expertise bridges clinical practice and data engineering.
Her career has centered on transforming real-world healthcare data—including EHR, FHIR, and claims datasets—into actionable clinical insights, focusing on trustworthy, explainable AI solutions for oncology and digital health.
Today’s oncology landscape is marked by a demand for actionable, real-time clinical support rooted in standardized data and clinical transparency. As hospitals race to integrate FHIR-based solutions and AI for cancer care, voices like Bodagala’s highlight both the promise and the essential safeguards required to move from back-office analytics to genuine bedside decision support.
Clinical context gaps
Fragmented information remains a persistent pain point for oncology clinicians when key data lives in disconnected systems. “The turning point was realizing how often oncology decisions were being made with incomplete longitudinal context. Key signals—treatment timing, toxicities, response patterns—were scattered across notes, labs, and orders that never came together in real time.”
Bodagala points out that while electronic dashboards help with reporting, “They didn’t change decisions.” The clinical imperative, she says, “FHIR-native, patient-level view that surfaced risk and trajectory at the moment of care, not after the fact.”
This scenario is reflected in recent implementations of robust oncology data models, such as the FHIR-based oncology data model in India, which incorporates dozens of oncology-specific resource profiles and standardized value sets to ensure semantic and syntactic interoperability across cancer care workflows.
Shaping data pipelines
Bodagala’s graduate training at the University of San Francisco and her FHIR data transformation work at Motive Medical Intelligence instilled an engineering ethos focused on workflow impact. “My graduate training emphasized that informatics succeeds or fails at the workflow boundary, not the algorithm.”
She notes that at Motive, her team prioritized treating interoperability as foundational. “Architecturally, that meant normalizing data into FHIR resources early, separating clinical semantics from modeling logic, and curating cohorts around real care pathways rather than billing artifacts.”
This perspective aligns with the evolution of open FHIR-based standards like mCODE, which codify minimal oncology data elements for diagnosis, genomics, treatment, and more, allowing for scalable, computable cancer research and evidence generation.
Data quality and standardization
Converting raw EHR streams into reliable clinical features is fraught with foundational challenges. Bodagala points to three recurring obstacles: “Inconsistent units, fragmented coding systems, and duplicated or poorly timed events.” These technical discrepancies “fundamentally distort clinical meaning.”
Her solution is to intervene early with canonical units, governed value sets, and clinically informed deduplication. She emphasizes: “If the semantics are wrong, no amount of modeling sophistication will compensate.”
The experience reflects broader operational barriers reported in deployments of automated extraction software like ExtractEHR, where local data mapping and value set standardization remain major hurdles in multi-institutional oncology research.
Sensitivity, specificity, and clinical trade-offs
AI deployment in oncology is inherently fraught with choices about how best to balance the risk of missing early warning signs versus minimizing false alarms. “In oncology, thresholds are clinical decisions. Missing early progression is far more costly than a false positive in many settings, but not all.”
Bodagala highlights the necessity of starting with calibration, then: “Tune sensitivity and specificity based on tumor type and care context.” She explains the clinical divergence: “A surveillance model and a palliative-care model should behave very differently—and clinicians immediately notice when those distinctions aren’t respected.”
This mirrors findings from evidence-gathering efforts using EHR-integrated, FHIR research platforms, such as prospective oncology data collection studies in community clinical practice, which emphasize the importance of data completeness for modulating model sensitivity and specificity.
Model transparency with SHAP
Model explainability is central for influencing clinician adoption. Bodagala recalls, “In one case, SHAP explanations highlighted treatment delays as stronger predictors than biomarkers I initially expected. That forced a re-evaluation.”
The effect was profound: “Clinicians recognized those delays as signals of toxicity management and care fragmentation—factors that weren’t always visible in traditional reviews. The explanation didn’t just justify the model; it changed how clinicians interpreted risk.”
Global frameworks, including the EBAI requirements for AI-based biomarkers, now emphasize clinical explainability, post-deployment monitoring, and ground-truth validation as minimum standards for adoption in cancer care AI.
Human-in-the-loop and explanation design
Bodagala stresses the importance of embedding AI decisions in ways that serve clinicians, not overwhelm them: “Predictions should support decisions clinicians are already making, not create new interruptions. We focus on passive surfacing—risk stratification, trend changes, and concise explanations embedded in the chart.” She observes: “Clinicians consistently prefer contextual explanations over pop-up alerts, especially when time pressure is high.”
This design principle is reinforced by expert models on human-in-the-loop oversight for clinical processes, which advocate for passive, context-aware information displays to foster trust and usability in the clinical workflow.
Validation, drift, and post-deployment rigor
The challenge of maintaining model robustness and trust after deployment is nontrivial. Bodagala employs a multi-layered validation approach: “We use temporal splits to avoid leakage, external testing where possible, and chart-review adjudication for edge cases.”
She adds, “Post-deployment, we monitor drift, subgroup performance, and silent failures—cases where outputs are ignored rather than wrong. In healthcare, loss of trust is as dangerous as loss of accuracy.”
This multifaceted approach parallels recent standards in FDA regulatory frameworks for AI/ML software in oncology, which require continuous performance monitoring, transparency, contextual validation, and risk-based oversight for adaptive models.
Adapting for emerging markets
Expanding explainable AI to under-resourced healthcare settings introduces different constraints. Bodagala adapts by simplifying technical and computational demands, while holding firm on data quality: “I would simplify aggressively—fewer features, lighter models—but never compromise on clinical semantics or temporal logic. FHIR alignment and data integrity are non-negotiable.”
Success metrics shift as well: “Impact in the first 90 days wouldn’t be measured by model metrics alone, but by earlier escalations, fewer missed follow-ups, and whether clinicians continue using the tool without being prompted.”
This reflects lessons from pilot research with FHIR-standardized oncology models and the adoption of modular digital systems like Personal Health Twin OS, which demonstrate how standardized interoperability facilitates adaptation across resource settings.
The emergence of explainable, FHIR-native clinical AI stands at a pivotal moment. Bodagala’s perspective underscores that technical sophistication alone is insufficient if clinical trust, transparency, and workflow integration are neglected. Through deliberate data engineering, contextual explanations, and rigorous validation, such systems can advance from pilot innovations to everyday clinical tools in oncology—delivering on the potential of AI while holding patient-centric safety as their guiding principle.