Sasi Kumar Kolla is a Healthcare AI and Machine Learning Systems Engineer whose work explores how intelligent systems can support better clinical care. His research brings together healthcare data platforms, predictive analytics, explainable AI, and real-time monitoring systems. Over the years, he has worked across areas such as electronic health record analytics, clinical risk prediction, federated learning, and multimodal healthcare data integration. He connects technical engineering with real clinical challenges, such as improving hospital readmission prediction models. His research explores areas such as generative AI, graph neural networks, and precision medicine, showing how modern AI tools may shape the future of healthcare delivery.

In this interview, Sasi Kumar Kolla shares thoughtful insights into the changing relationship between healthcare infrastructure and artificial intelligence. He also discusses the challenges of clinical AI adoption, the importance of transparency in decision-making systems, and the future of autonomous healthcare technologies in an increasingly connected medical world.

Q1. Sasi, thank you for joining us today. Your work spans from foundational EHR architectures to advanced AI-driven clinical systems. How do you see the relationship between data infrastructure and intelligent clinical insights evolving as healthcare systems become increasingly real-time and interconnected?

Sasi Kumar Kolla: The relationship between data infrastructure and clinical intelligence has always been symbiotic, but what’s changing now is the directionality of that relationship. Historically, infrastructure was built first, and intelligence was layered on top, often as an afterthought. You had EHR systems designed primarily for documentation and billing, with analytics teams trying to extract meaning from data structures that were never intended for real-time decision support. The insights were inherently delayed, retrospective, and limited by what the infrastructure could surface.

What I see evolving and what my own research trajectory has been oriented toward is a fundamental inversion of that model. In truly interconnected, real-time healthcare systems, intelligence needs to be designed into the infrastructure from the ground up, not appended to it. The pipeline architecture, the data contracts between systems, the latency tolerances, and the schema decisions all have to be made with the downstream clinical question in mind. When you’re building a real-time patient deterioration prediction system using deep learning, as I’ve been working on recently, you cannot retrofit that capability onto a batch-oriented EHR platform. The infrastructure has to be engineered to deliver the right signals at the right granularity, in near-real time, reliably.

The second evolution I find compelling is multimodal data integration, the convergence of EHR records, imaging data, genomic profiles, and even wearable or remote monitoring streams into a unified clinical picture. Each of those data types has its own infrastructure requirements, latency profile, and quality challenges. Building the connective tissue between them, the pipelines, the normalization layers, and the federated governance models is where the most important and underappreciated engineering work in healthcare AI is happening right now. The clinical insights are only as intelligent as the infrastructure allows them to be.

Q2. In your earlier research, you spent time on creating secure and interoperable data pipelines for multi-hospital environments. Looking back, what were some lesser-discussed technical or systemic limitations in those early frameworks that pushed you toward more advanced AI-driven approaches later on?

Sasi Kumar Kolla: My early research on large-scale EHR data platforms and secure clinical data exchange gave me a deep appreciation for what those frameworks could accomplish and an equally deep understanding of where they fundamentally broke down.

The most discussed limitation in the literature was interoperability, HL7, FHIR, the challenge of getting disparate hospital systems to speak the same data language. That problem is real, but it’s also the one that received the most engineering attention and funding. The less-discussed limitations were more subtle and, in some ways, more consequential.

The first was semantic inconsistency beneath the surface of syntactic compliance. Two hospitals could exchange data using the same FHIR schema and still produce clinically incomparable records, because the coding practices, documentation habits, and clinical vocabularies of their providers differed systematically. A “high-risk” annotation in one system meant something meaningfully different in another. Early pipelines treated data exchange as a solved problem the moment the format was standardized, but the semantic layer, which is what the data actually means clinically, remained fragmented and ungoverned.

The second was static rule brittleness. Early clinical decision support systems were built on expert-defined rules: if creatinine exceeds X, flag for renal risk. Those rules encoded the knowledge of the clinicians who wrote them at a specific point in time, for a specific patient population. They didn’t adapt. They didn’t learn from outcomes. And when patient populations shifted, or new clinical evidence emerged, the rules quietly became wrong. That brittleness is what drove me toward machine learning frameworks for clinical risk prediction, not because ML is inherently superior, but because it can update its understanding from data in ways that static rule systems cannot.

The third limitation was privacy as a barrier to learning. The most valuable thing about multi-hospital data is the scale and diversity it provides for training robust models. But early interoperability frameworks had no mechanism to enable learning across institutions without centralizing data, which created an irresolvable tension with HIPAA and patient privacy expectations. That tension is what pushed me toward federated learning as a research focus, building systems that can learn from distributed data without the data ever leaving its originating institution.

Q3. In your work on predictive modeling for hospital readmission and population health analytics, you explored proactive healthcare strategies. How do you balance the tension between predictive foresight and clinical practicality, especially in environments where resources and response capacities may be limited?

Sasi Kumar Kolla: This tension is one of the most important and most underappreciated design constraints in healthcare AI, and I’ll be honest: it’s something I wrestled with significantly in my work on predictive modeling for hospital readmission risk and population health analytics.

The naive framing of predictive healthcare is that more prediction is always better. If your model can identify a thousand high-risk patients, you’ve done your job. But in a resource-constrained clinical environment, which is most clinical environments, a list of a thousand high-risk patients is not an insight. It’s a burden. If the care coordination team can meaningfully follow up with fifty patients in a given week, a model that surfaces a thousand is generating noise, not value. Worse, it can erode clinical trust in the system entirely, because the actionable signal gets lost in the volume.

My approach to this tension has been to design predictive systems around response capacity, not just risk probability. This means working closely with clinical operations teams before model deployment to understand what intervention resources actually exist, what triggers a realistic clinical response, and what the cost of a false positive versus a false negative looks like in that specific setting. The model outputs are then calibrated to that operational reality, not just optimized for AUC in isolation.

The second piece is tiered prediction horizons. Not all predictions need to be acted on immediately, and designing for that distinction matters enormously. A 72-hour readmission risk flag requires a different response pathway than a 30-day population health risk score. When predictions are stratified by urgency and actionability, resource allocation becomes more tractable, and clinical teams can engage with the system’s outputs without feeling overwhelmed.

What I’ve found is that the models that have the most real-world clinical impact are rarely the most statistically sophisticated. They’re the ones designed in genuine partnership with the clinicians and care teams who will use them, built around the rhythms and constraints of actual care delivery, not just theoretical optimization metrics.

Q4. You’ve contributed to the development of explainable AI systems and bias mitigation strategies in clinical decision-making. In your view, what does “meaningful transparency” actually look like for clinicians using these systems, and how can it go beyond just interpretability into something actionable in real-world care settings?

Sasi Kumar Kolla: This is a question I care about deeply, and my research on explainable AI and bias mitigation in clinical decision-making has shaped a strong perspective on it.

“Interpretability” in the current AI discourse tends to mean one thing: being able to explain why a model made a specific prediction. SHAP values, LIME scores, and feature importance rankings are the tools most commonly offered to satisfy the transparency requirement. And they have real value. But for a clinician standing at a bedside making a time-sensitive decision, a ranked list of contributing features is not transparency. It’s a technical artifact that may or may not map to anything clinically meaningful.

Meaningful transparency, as I define it, has three properties that go beyond interpretability. The first is clinical contextualization; the explanation has to be expressed in the clinician’s language and clinical frame of reference, not the data scientist’s. Instead of “feature X contributed 0.34 to the risk score,” meaningful transparency says “this patient’s risk is elevated primarily because of a pattern of deteriorating kidney function indicators over the past 72 hours, combined with a history of prior readmission.” That’s actionable clinical reasoning, not a statistical footnote.

The second property is uncertainty communication. Most AI systems present a point estimate of a risk score, a probability without conveying the confidence interval around that estimate or the conditions under which the model’s reliability degrades. A clinician should know whether a risk score of 0.78 is coming from a well-calibrated region of the model’s training distribution or from an edge case where the model is essentially extrapolating. That distinction matters enormously for how much weight to place on the prediction.

The third property is bias transparency, and this is where my bias mitigation research becomes directly relevant. Meaningful transparency includes surfacing whether a model has known performance disparities across demographic subgroups. If a readmission risk model performs well for the majority population in the training data but underperforms for elderly patients or certain ethnic groups, the clinician using it for those patients deserves to know that. Hiding that information in the name of a clean user interface is a form of opacity that can cause direct clinical harm.

Transparency that encompasses all three of these properties doesn’t just help clinicians use AI systems better; it builds the kind of calibrated trust that allows AI to augment clinical judgment rather than replace or genuinely mislead it.

Q5. Your recent work explores generative AI for synthetic healthcare data and graph neural networks for drug interaction analysis. Could you walk us through a specific use case where combining these technologies has the potential to significantly change how medical research or treatment planning is conducted?

Sasi Kumar Kolla: Let me walk through a concrete scenario that I think illustrates the transformative potential of combining these two technologies.

Consider the challenge of rare disease drug interaction research. For a condition affecting fewer than 200,000 patients in the United States, the real-world clinical data available for research are severely limited by definition. Traditional pharmacological studies require sufficient sample sizes to detect interaction signals with statistical confidence, and for rare diseases, that threshold is often simply unachievable with observed patient data. The consequence is that rare disease patients are treated with incomplete pharmacological knowledge, and clinicians are making drug combination decisions based on extrapolation from more common conditions that may or may not be physiologically analogous.

This is where the combination of generative AI and graph neural networks becomes genuinely powerful. Generative AI, specifically diffusion models or transformer-based architectures trained on broader pharmacological and genomic datasets, can produce synthetic patient profiles that are statistically plausible representations of rare-disease patients, preserving the distributional characteristics of the condition without exposing any real patient data. This synthetic cohort can be generated at whatever scale the research question requires, solving the sample size problem that makes rare disease research so difficult.

Those synthetic patient profiles can then be fed into a graph neural network designed to model drug interactions, in which nodes represent drugs, metabolic pathways, and molecular targets, and edges represent known or probabilistically inferred interactions. GNNs are particularly well-suited for this because drug interactions are inherently relational; a drug doesn’t just have properties in isolation; its behavior is a function of how it sits within a network of molecular relationships. The GNN can propagate information through that network to predict interaction effects for drug combinations that have never been clinically observed, flagging high-risk combinations for avoidance and identifying potentially synergistic combinations for further investigation.

In treatment planning terms, a system like this could allow an oncologist treating a patient with a rare tumor type to receive a ranked assessment of potential drug combinations, with predicted interaction profiles, calibrated to the specific molecular characteristics of that patient’s profile, precision medicine at a level that current research data volumes simply cannot support. That’s a qualitative change in what’s possible, not just a quantitative improvement on existing methods.

Q6. Moving forward, as you integrate large language models, genomics, imaging, and EHR data into autonomous clinical monitoring systems, what new kinds of interdisciplinary collaboration do you believe will become essential? How might this reshape the traditional boundaries between data scientists, clinicians, and healthcare institutions?

Sasi Kumar Kolla: The integration of large language models, genomics, imaging, and EHR data into autonomous clinical monitoring systems is not an engineering problem that any single discipline can solve on its own. And I think the interdisciplinary collaboration it demands will be genuinely new, not just more of the existing collaborations between data scientists and clinicians, but structurally different in ways that will challenge traditional professional boundaries.

The first new collaboration that I believe becomes essential is between data scientists and clinical ethicists, working together as genuine equals in system design, not with ethicists reviewing finished systems for compliance issues, but embedded in the design process from the beginning. When an autonomous monitoring system is making real-time decisions about patient deterioration alerts, the ethical questions about false-alarm rates, how uncertainty is communicated, and whose values are encoded in the alert thresholds are not separable from the technical questions. They have to be resolved together, which requires an institutional structure that places ethical reasoning and technical reasoning in genuine dialogue.

The second collaboration is between genomicists and ML engineers at the systems architecture level. Genomic data have properties that general-purpose ML infrastructure handles poorly: extreme dimensionality, complex linkage disequilibrium structures, and population stratification. The collaboration needed is not “here is the data, build a model,” but joint design of the data representation, the model architecture, and the validation approach. That requires genomicists willing to engage with computational constraints and ML engineers willing to engage with biological reality.

The third shift, and perhaps the most institutionally significant, is the one happening at the boundary between healthcare institutions and AI research organizations. For multimodal autonomous systems to work safely at scale, they need to be trained, validated, and monitored on data from diverse patient populations across multiple institutions. That requires data-sharing agreements, federated-learning infrastructure, and governance frameworks that don’t currently exist in a mature form. Building them will require institutional leadership, legal, clinical, and technical collaboration in ways that healthcare systems and research organizations are not traditionally structured to support.

What I anticipate is that the most successful healthcare AI deployments of the next decade will not be defined by the sophistication of their models. They will be defined by the quality of the institutional infrastructure and the interdisciplinary trust that surrounds those models, as well as the ability to bring clinicians, data scientists, ethicists, genomicists, and institutional leaders into a shared practice of responsible, iterative, evidence-based AI deployment. That infrastructure is the hard problem. And building it is the work I find most meaningful and most urgent.

Conclusion

Sasi Kumar Kolla offers a thoughtful look into the future of healthcare AI and the systems shaping modern clinical care. Progress in healthcare technology is not only about building smarter models, but about creating systems that clinicians can trust and patients can benefit from in meaningful ways. His work highlights the growing need for intelligent systems that remain practical, transparent, and grounded in real clinical settings. Sasi also emphasizes the importance of strong data foundations, collaborative system design, and ethical responsibility. His views on federated learning, bias transparency, and real-time clinical monitoring point toward a healthcare future influenced by technical depth and human understanding.