Opinions expressed by Digital Journal contributors are their own.
Emergency departments across the United States process cardiac cases where minutes determine outcomes. Electrocardiogram machines record cardiac activity in seconds, but getting those readings interpreted can take hours, especially in hospitals without cardiologists on staff. The holdup isn’t the technology capturing the data. It’s what happens next: turning waveform patterns into reports that emergency physicians can use to make treatment decisions. Many facilities send ECG readings to off-site specialists who review them remotely, and patients wait.
Healthcare data volumes have grown faster than the infrastructure to interpret them. Over 300 million ECGs are performed annually in the United States alone, yet hospitals generate terabytes of imaging, diagnostic readings, and transaction records daily, with much of this information sitting in queues waiting for human analysis. Manual processes create variability; two clinicians reviewing identical ECG traces may reach different conclusions, particularly in borderline cases. The challenge extends beyond cardiology into dental imaging, where wisdom tooth assessments depend on radiologist availability, and healthcare billing, where fraudulent claims slip through retrospective audits that flag problems months after payments are clear.
Traditional machine learning excels at sorting data into categories but struggles when asked to explain findings in language that clinicians understand or adapt when operational conditions shift. A model trained on one hospital’s patient population may underperform in another. Fraud detection rules become obsolete as billing schemes evolve. These gaps have pushed researchers toward systems that don’t just recognize patterns but generate explanations and learn from production environments without constant retraining.
Research that generates clinical narratives, not just classifications
Srinubabu Kilaru has built AI architectures for settings where accuracy matters as much as speed. His work addresses diagnostic delays in cardiology, dental imaging, financial fraud, and autonomous robotics, domains where decisions carry clinical or operational consequences and where traditional models often fail when real-world conditions diverge from training data.
His RAEMAS 2025 publication introduced a system for automated ECG report generation using retrieval-augmented generation. The architecture combines one-dimensional convolutional neural network encoders with ECG-BERT-based retrieval and reinforcement learning. The system doesn’t just flag irregular heartbeats or potential problems. It writes out full reports, the kind a cardiologist would normally produce after reviewing the same data. When tested against standard benchmarks, the model scored 41.3 on BLEU, 53.5 on ROUGE-L, and 38.7 on METEOR, measurements that compare generated text against human-written examples. Those numbers put the output close to what physicians write themselves.
Most medical AI stops at identifying patterns and assigning probability scores. Doctors still have to interpret those results and write up findings in clinical language. His approach skips that step, and his system retrieves relevant interpretations from validated ECG databases and uses reinforcement learning to refine phrasing based on clinical accuracy. The framework has been extended to EEG analysis and multilingual diagnostic contexts, where language barriers compound interpretation delays.
“The objective is compressing time between data capture and usable clinical insight, particularly where specialists aren’t immediately available,” Srinubabu Kilaru explained. “These systems provide frontline clinicians with initial interpretations while definitive reviews are arranged.”
At ICCCNT 2025, held at IIT Indore, he presented two additional diagnostic systems. The first applies convolutional architectures to wisdom tooth detection in dental X-rays, a task complicated by anatomical variation and inconsistent imaging quality across different equipment and patient ages. The second details temporal graph neural networks for real-time fraud detection in financial transaction streams, designed to identify anomalies as transactions occur rather than in post-processing reviews.
Fraud detection that operates in transaction time
Healthcare billing fraud represents a persistent drain on federal programs. Traditional fraud analytics flag suspicious activity through batch processing, and auditors review claims weeks or months after payments have been issued, limiting recovery options.
Kilaru’s temporal graph neural network models transactions as dynamic networks where providers, patients, and services form nodes, and their relationships create edges that evolve. The system identifies structural anomalies in near-real time by learning normal transaction patterns and flagging deviations before payment processing completes. This approach addresses fraud schemes that evolve deliberately to evade static rule sets, adapting to new patterns without waiting for labelled examples of emerging attack types.
The architecture has applications beyond billing fraud. Similar graph-based methods can track supply chain anomalies in hospital inventory systems or identify irregular prescribing patterns in pharmaceutical distribution networks. The underlying principle, modelling relationships that change dynamically rather than analyzing snapshots, applies wherever temporal context matters for detection.
Production systems that handle millions of eligibility determinations
In June 2023, eSystems Inc. gave him a Gold Award for his work leading the architecture on a state Medicaid eligibility system. The platform handles verification checks, enrollment processing, and data sharing for millions of people who rely on the program. Building it meant pulling information from dozens of separate databases, federal tax records, state citizenship files, healthcare provider registries and making sure caseworkers got answers in under a second when they looked up an applicant.
The system runs on microservices that operate independently, processes data as events happen rather than in batches, and caches information strategically to avoid repeat lookups. Other states have used the same architectural model when building their own Medicaid platforms. During open enrollment windows, when applications jump by ten times normal volume within a few days, the system scales up based on past patterns and current traffic rather than just throwing more servers at the problem.
Kilaru’s work on CGI’s APEX program, which earned him a Pegasus Award, dealt with similar issues, taking older government systems and making them handle enrollment surges without crashing or slowing down. The focus involved refactoring legacy systems to manage enrollment surges without degrading performance, a problem common across public benefit programs that operate under strict processing deadlines and regulatory oversight.
Adaptive control for autonomous hospital systems
Kilaru’s 2013 research in the International Journal of Intelligent Engineering and Systems examined adaptive fuzzy control systems for autonomous mobile robots, focusing on energy optimization. The controllers adjusted motor output based on terrain, obstacle density, and remaining battery power, extending operational range in hospital logistics and warehouse environments.
Healthcare facilities have increased deployment of autonomous systems for medication delivery, supply transport, and UV disinfection. These environments differ from structured industrial settings—hospital layouts change constantly, human traffic is unpredictable, and safety margins are zero. The adaptive control methods he developed allow robots to learn operational environments over time rather than requiring exhaustive pre-mapping, a critical capability where corridors are reconfigured and foot traffic patterns shift by the hour.
His collaboration with Dr S. Satyanarayana at Malla Reddy University connects academic research with deployment challenges in healthcare AI, fraud analytics, and automated diagnostics. He also serves as a peer reviewer for several journals and has participated as a guest mentor and panellist, addressing gaps between research prototypes and production-ready systems. In 2025, he was selected as a judge for the Business Intelligence Group’s Sustainability Award, evaluating projects combining analytics, sustainability, and technology.
Systems that adapt without complete retraining cycles
The technical shift toward retrieval-augmented generation and reinforcement learning reflects recognition that static models degrade in production. Model drift, the gradual decline in accuracy as real-world data diverges from training conditions, affects diagnostic systems deployed across different patient demographics or imaging equipment. A cardiac model trained on data from one hospital may underperform elsewhere. Fraud detection rules become obsolete as schemes evolve.
Retrieval-augmented systems address this by referencing updated knowledge bases and adjusting to local performance benchmarks without full retraining. Temporal graph architectures in fraud detection identify novel patterns without waiting for labelled examples of new attack types. These approaches acknowledge that deployment environments change and that models need mechanisms to respond without returning to training pipelines.
Diagnostic AI moving from pilot projects to operational infrastructure must meet criteria beyond test accuracy: outputs need to be explainable, systems must adapt to local conditions, and performance must meet real-time constraints. Srinubabu Kilarus’s work across ECG report generation, billing fraud detection, and adaptive robotics demonstrates that clinical-grade AI depends on reliability when variables shift and when decisions carry consequences, not just performance in controlled validation tests.
