KBKiran Bhattacharyya

Kiran Bhattacharyya

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Kiran Bhattacharyya has a total of 15 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    The economics of accuracy for medical reasoning with large language models

    by Kiran Bhattacharyya, Sreeram Kamabattula

    Deploying large language models (LLMs) in clinical settings is limited by security, reliability, latency, and accessibility concerns that favor smaller, on-device or on-premise models. However, these smaller models may struggle to meet accuracy requirements. While fine-tuning and retrieval-augmented generation (RAG) can improve domain-specific accuracy, these methods require additional labeled data, technical skill, and infrastructure. In contrast, test-time scaling —allocating extra token-budget during inference—offers a training-free alternative to increasing accuracy. However, the trade-offs between these strategies and their interaction with model size remain poorly understood for medical reasoning. To address this gap, we compare three approaches—test-time scaling, fine-tuning, and context grounding—using the Gemma and MedGemma family of LLMs (Gemma-3 1B, Gemma-3 4B, Gemma-3 27B, MedGemma-4B, and MedGemma 27B) and evaluate these systems across common biomedical question-answering (QA) datasets and a set of recently released medical exam questions with the performance of practicing clinicians available for comparison. We test baseline prompts (direct answer, Chain-of-Thought, and self-consistency) while introducing a new prompting method we call “prompt-chaining for continuous reflection” (PCCR) that forces inference time minimum token-generation budgets. We assess accuracy and tokens-generated, allowing us to investigate the accuracy–efficiency trade-offs across prompting, context-grounding, fine-tuning, and model scales. We discover equivalency point configurations where a smaller model’s accuracy falls within one 95% confidence interval of a larger model’s (typically within 1–4 percentage points) reached through increased reasoning budgets, context-grounding, or fine-tuning. Specific effects are apparent and statistically supported: the benefit of medical fine-tuning grew from +4.6 to +15.7 percentage points (non-overlapping 95% CIs) when paired with self-consistency, and enforced extended reasoning raised MedGemma 27B accuracy from 58.1% to 80.1% ( p < 10 -5 ) in the absence of context. We also identify an “overthinking” inflection: when high-quality context is available, extended reasoning beyond roughly 128–256 tokens degrades accuracy by 7–13 percentage points. Using these empirical results, we formulate a general framework with equations to balance cost-benefit trade-offs when engineering LLM-based systems for medical reasoning and QA. We recommend generalizable configurations, designs, and patterns to achieve accuracy and efficiency objectives for example use-cases relevant to healthcare organizations.

  • Open Access

    Multi-fidelity Bayesian optimization of population-robust near-infrared sensors for skeletal muscle oximetry

    by Kiran Bhattacharyya

    Designing transcutaneous skeletal muscle oxygenation (SmO 2 ) sensors requires jointly optimizing source–detector geometry and wavelength selection while guaranteeing performance across populations that vary in subcutaneous fat thickness and skin pigmentation. We present a multi-fidelity Bayesian optimization (MFBO) framework that couples Monte Carlo light-transport simulations at two photon-count fidelities to a distributionally robust design objective. An autoregressive Gaussian-process surrogate learns the correlation between inexpensive low-photon-count and accurate high-photon-count simulations, and a cost-aware acquisition function decides both where and at what fidelity to sample. Robustness across the population is enforced with Conditional Value-at-Risk (CVaR) and entropic-risk (ERM) objectives that target worst-case subjects rather than the population average. On a five-layer forearm tissue model with anthropometric variability we find (i) a fidelity regime that is favorable for MFBO where the low-fidelity surrogate is rank-informative (Spearman ρ = 0.84) but biased, at 100× lower cost; (ii) MFBO attains 23% higher robust sensitivity than a strong high-fidelity single-fidelity baseline at equal budget ( p = 0.035), and avoids the optimistic bias that causes low-fidelity-only optimization to collapse when its designs are validated at high fidelity; (iii) CVaR/ERM objectives improve worst-case tail performance by ~ 23% relative to a mean objective without sacrificing average sensitivity; and (iv) discovered designs improve robust tail sensitivity by roughly 3–6× over commercial and heuristic optode layouts, with the largest gains in the high-fat and high-melanin subpopulations. The methodology bridges stochastic light-transport physics with sample-efficient machine-learning optimization and generalizes to cerebral oximetry, photodynamic therapy planning, and wearable physiological monitors.


  • Continuous tracking of aortic aneurysm diameter with peripheral pulse waves: a computational framework combining sequential Markov chain Monte Carlo with Kalman filtering.

    by Kiran Bhattacharyya

    Abdominal aortic aneurysms (AAA) affect more than 1% of adults over 50 and carry significant mortality risk. Current surveillance relies on intermittent imaging (ultrasound or MRI) at 6--24 month intervals, which may miss rapid growth acceleration between visits. We investigate the feasibility of continuous aneurysm diameter tracking using peripheral pulse waves, like those detected by photoplethysmography (PPG) devices. Approach. We use a simplified one-dimensional (1D) hemodynamic model that simulates pulse wave propagation from the heart to the pedal digital artery. We first demonstrate diameter estimation when the hemodynamic model parameters defining systemic circulation are known within bounds for an individual, aggregating thousands of observations over hours or days. We then address the more challenging scenario where systemic parameters are only known to be within wider population-level bounds, using a sequential Monte Carlo approach that combines an ensemble Markov chain Monte Carlo (MCMC) sampler with Kalman filtering to marginalise over unknown parameters while tracking the aneurysm diameter. Both approaches are evaluated through 12-month tracking simulations of virtual patients with constant and accelerating aneurysm growth. Main results. While single-observation diameter estimation is fundamentally limited by noise and confounding variables, aggregating 1,600 measurements under baseline noise conditions reduces diameter uncertainty to 0.8 mm when patient-specific hemodynamic parameters are known within bounds. In this setting, tracking simulations across eight virtual patients achieve average root-mean-square error (RMSE) of ∼0.3 mm. When systemic parameters are known only within population-level bounds, joint Bayesian estimation over the full parameter space achieves a median RMSE of 0.65 mm (1.4±0.3 mm, mean±standard error) across 50 virtual patients, remaining within clinically relevant ranges despite the underlying parameters being only partially identifiable. Significance. These computational results suggest that peripheral pulse wave monitoring through PPG sensors could complement traditional imaging for AAA, potentially enabling earlier detection of growth and more timely intervention.

  • Learning where to look: scaling parkland grade prediction from surgical videos

    by Sreeram Kamabattula, Sue Kulason, Busisiwe Mlambo, Lilia Purvis, Kiran Bhattacharyya

    Purpose: The Parkland Grading Scale (PGS) is widely used to quantify operative difficulty in cholecystectomy, with higher grades associated with worse post-operative outcomes. However, consistent, scalable PGS assessment is limited by the reliance on two manual steps: determining where to look in the surgical video for key evidence, and assigning a grade. Previous machine learning approaches have either depended on manual selection of where to look, or approximated it with fixed-duration video segments, leaving it unclear whether models can accurately predict PGS without explicit guidance on where to look. Methods: To address this, we evaluate 287 robotic cholecystectomy videos annotated with PGS and a standardized key-segment. Using a temporal convolution network and attention-based framework, we compare the performance of a fully automated model using full surgical videos without key-segment supervision to a model provided with the key-segment (where to look). Results: Providing the key-segment yields substantial performance gains (weighted F1 +0.25 and Krippendorff’s \(\alpha \) (KA) +0.29). We further introduce ParkNet\(_{LEARN}\), which learns to where to look and predicts PGS from full surgical videos, achieving significant improvements over the no-supervision automation (weighted F1 +0.18 and KA +0.23), and a KA = 0.60–within 0.06 of the model with key-segment provided. Conclusion: These findings highlight the importance of attending to where to look for automating operative difficulty assessment, and is a valuable step toward supporting large-scale research on surgical performance and post-operative outcomes.

  • A Bayesian approach to temporal surgical segmentation model fusion

    by Max Berniker, Sreeram Kamabattula, Kiran Bhattacharyya

    Purpose: Robotic-assisted surgery (RAS) generates vast amounts of video and robotic data, presenting opportunities for machine learning. Video-based models, in particular, that can temporally segment frames by ontological categories such as procedure type, phase, steps, actions, etc., are needed. Training separate models for each category neglects statistical dependencies between categories and can yield incompatible predictions. Training large multi-category models may help, but increases complexity while reducing model modularity and interpretability. Methods: We present a model fusion alternative: an effectively zero-free-parameter Bayesian model fusion technique. Incorporating the empirical conditional dependencies across categories and time, we combine predictions from multiple segmentation models into one joint Bayesian inference. The result is a Bayes’ optimal distribution over all categories evolving over time with accumulated evidence. Results: On a large test set of hundreds of surgical cases, of nearly eight million frames of annotated data, we found that fused predictions from the joint Bayesian model provide clear benefits over the individual models, correcting inconsistent and inaccurate predictions, and even forming accurate beliefs when evidence was absent. Conclusion: The model we present is a lightweight, principled alternative to machine learning-based model fusion. A sufficiently complex model could be trained to produce the same results, but would effectively trade explainable predictions with minimal overheard for computational complexity and transparency. We end by discussing how the same approach can be used to encompass larger more sophisticated models within the same conceptual framework.

  • Open Access

    Longitudinal Tracking of Valve Regurgitation from Peripheral Photoplethysmography Using a 1D Hemodynamic Surrogate: a Joint MCMC-Kalman Framework

    by Kiran Bhattacharyya

    A bstract: Aortic (AR) and mitral (MR) regurgitation are progressive valvulopathies whose management depends on tracking severity over months to years, currently via infrequent clinic-bound echocardiography. We investigated whether peripheral photoplethysmography (PPG) can track AR and MR longitudinally despite unknown day-to-day hemodynamic variation. A one-dimensional Navier–Stokes arterial model from aortic root to pedal digital artery, with parametric AR and MR at the inlet, generated distal waveforms from which we extracted eleven morphological features. Univariate sensitivity, Cramér– Rao bound, and feature-ablation analyses quantified theoretical observability under realistic noise and physiological variability. A neural surrogate of the 1D model was embedded in a joint Markov Chain Monte Carlo (MCMC)–Kalman scheme that tracked regurgitant fraction monthly while marginalizing over unknown systemic parameters. Across 240 synthetic patients spanning stable and high-risk AR and MR, with and without progressive hypertension, over 36 months, the tracker achieved RMSE of 0.023–0.028 for AR and 0.075–0.101 for MR in regurgitant-fraction units. Dicrotic notch timing was most informative for both valves; augmentation index was uniquely critical for MR. Progressive hypertension paradoxically improved MR tracking by 16–18%. MCMC-Kalman tracking errors arose mainly from over-constrained latent parameters. These results support the theoretical feasibility of wearable PPG-based surveillance of valve regurgitation and motivate prospective clinical validation.

  • Open Access

    Digital Twins for Fungal Computing: Viable XOR Regimes, Parameter Inference, and Waveform-Guided Rediscovery

    by Kiran Bhattacharyya

    Fungal substrates are promising candidates for unconventional computing, but specimen-to-specimen variability makes logic-gate fabrication difficult to reproduce. This paper presents a digital-twin workflow for fungal excitable networks and evaluates three components needed for computeraided design: identifying parameter regimes that support XOR computation, inferring latent biophysical parameters from electrical characterization data, and refining those inferred parameters by waveform matching. The model represents mycelium as a random geometric graph with FitzHugh–Nagumo node dynamics and memristive edge conductances. A systematic optimization study over 160 simulated specimens identifies a viable XOR subspace defined by tuned biophysical parameters, electrode geometry, and stimulus timing. A characterization study over 400 simulated specimens uses step-response, paired-pulse, and triangle-sweep protocols to extract 94 response features. Random forest regressors recover several latent parameters reliably ( R 2 = 0.912 for τ v , 0.816 for τ w , 0.717 for a ), while v scale , R on , and R off remain weakly identifiable. On a preliminary rediscovery validation using 15 optimized specimens (20–50 nodes), ML initialization followed by local waveform-matching refinement reduces mean waveform mismatch from 1.070 to 0.042 (96.0%; one-sided Wilcoxon p = 3.1 × 10 −5 ) and reduces mean core-parameter error from 16.6% to 8.8% ( p = 6.1 × 10 −5 ). A sensitivity analysis on 72 viable specimens reveals that τ w and α are the most consequential parameters for XOR twin accuracy, while v scale and R off are both hard to identify and tolerant to error. These results show that fungal digital twins can already narrow the search for viable computational substrates, partially recover the excitable dynamics that govern them, and support small-scale specimen-specific refinement without yet claiming full XOR transfer.

  • Surgical Activity Recognition with Frame-Action Cross-Attention Temporal Modeling

    by Sreeram Kamabattula, Conor Perreault, Kiran Bhattacharyya

    Accurate surgical activity recognition from surgical video can support intraoperative decision-making, post-operative analysis, and surgical education. However, complex and variable workflows across different procedures challenge the generalization of current models. Consistently learning temporal dependencies at the frame-level, activity-level, and case-level across different procedure types remains an open problem. Here, we apply the existing Frame-Action Cross-Attention for Temporal modeling (FACT) architecture to the surgical domain, since it was designed for this purpose. We evaluate its performance across three datasets: Cholec80 (cholecystectomy, 7-phase), AutoLaparo (hysterectomy, 7-phase), and MultiBypass140 (gastric bypass, 12-phase, and 46-step). To distinguish between the relative importance of spatial and temporal representation learning, we investigate both, domain-general and surgically fine-tuned image feature extractors with FACT which jointly reasons over frame-level and action-level dependencies across the entire case via bidirectional cross-attention. To better understand workflow variability in our datasets, we model the variance in surgical workflows across these datasets using metrics. We compare FACT, which processes an entire case as a single sample, against multiple other architectures, including traditional windowed methods to evaluate how different approaches perform across these datasets. Across Cholec80, AutoLaparo, and MultiBypass140, FACT delivers competitive performance under matched protocols including 94.1% accuracy on Cholec80 phases, 77.3% on MultiBypass140 steps, and 89.5% accuracy on MultiBypass140 phases. Importantly, FACT performs comparably when using either domain-general or surgically-finetuned spatial representations, in some cases, even being relatively robust to older generations of domain-general image feature extractors, like RotNet. We emphasize method clarity, cross-dataset consistency, and workflow-variability analysis as our key contributions. While our results suggest that FACT is a flexible and robust temporal architecture for surgical activity modeling, challenges remain in generalizing to fine-grained or highly variable workflows. We also include a discussion of how the performance of different architectures may be affected by the measured variance in surgical workflows and propose future work that may help close this gap.


  • Open Access

    Decoupling Reasoning and Reward: A Modular Approach for Stable Alignment of Small Clinical Language Models

    by Kiran Bhattacharyya, Sreeram Kamabattula

    Deploying language models (LMs) in clinical settings requires navigating competing demands between accuracy, auditability, and on-device efficiency for privacy. While smaller LMs are desirable for this purpose, aligning them with methods like Group Relative Policy Optimization (GRPO) is often hindered by training instability and objective conflicts. Prior work has shown that Chain-of-Thought (CoT) supervision during supervised fine-tuning (SFT) can stabilize GRPO, but existing approaches typically entangle these objectives within a single, monolithic model. In this work, we introduce a modular, adapter-based alignment framework that decouples reasoning supervision and reward tuning into separate, composable parameter-efficient fine-tuning (PEFT) stages using LoRA adapters. We evaluate five alignment configurations on a medically grounded question answering dataset, using Qwen2.5 models from 0.5B to 7B to analyze how alignment stability, factual accuracy, and structural auditability scale with model size. Our findings demonstrate that this modular approach resolves key training instabilities, especially in smaller models, and produces structurally consistent, auditable reasoning without sacrificing accuracy. To support further research, we release, upon publication, (1) our dataset comprising over 100K clinically relevant QA pairs with CoT traces and (2) our multi-stage alignment codebase. We conclude that decoupling reasoning and reward offers a flexible and robust foundation for building privacy-preserving, verifiably aligned clinical LLMs that successfully address the competing demands in clinical AI.