About

Petar M. Djuric is a Distinguished Professor at Stony Brook University, specializing in Machine Learning, Artificial Intelligence, Signal and Information Processing, Bayesian Optimization, Monte Carlo methods, and Distributed Signal Processing. His research focuses on developing advanced algorithms and techniques in these areas, contributing to the understanding and application of signal processing and machine learning in various complex systems. His work is recognized for its impact on the theoretical foundations and practical implementations of intelligent systems, aiming to enhance data analysis, decision-making processes, and computational efficiency in engineering and scientific domains.

Research topics

• Computer Science
• Artificial Intelligence
• Machine Learning
• Clinical psychology
• Medicine
• Algorithm
• Telecommunications
• Psychiatry
• Mathematics
• Optics
• Psychology
• Geography
• Physics
• Family medicine
• Statistics
• Real-time computing

Selected publications

• State Space Clustering for Interpretable Fetal Heart Rate Characterization

  2026-04-21

  articleOpen accessSenior author

  Computerized cardiotocography (CTG) often relies on fetal heart rate (FHR) features with weak correlations to umbilical cord blood pH, the gold standard for neonatal acidosis, whereas deep learning features lack interpretability. In this work, we propose a novel family of interpretable FHR features based on clustering state vectors from shadow manifolds reconstructed from FHR signals. This approach partitions recordings into low- and high-volatility regimes and derives features that quantify both cluster compactness and separation. Experiments on an open-access CTG database demonstrate that these features achieve substantially stronger correlations with cord blood pH than conventional variability and nonlinear measures, which indicates their potential to improve the diagnostic power of computerized FHR analysis.

  PublisherOA PDFDOI

• Continual Time Series Forecasting with Diffusion Models Under Functional Regularization

  2026-04-21

  articleSenior author

  Diffusion models have recently achieved state-of-the-art performance on time series inference tasks such as forecasting and imputation. However, these models are usually trained unconditionally in a dataset-specific manner, which limits their ability to be applied cross-domain. Sequentially training the model on new tasks often causes a loss of predictive capability on previously encountered tasks. To address this challenge, we propose a functional regularization strategy that penalizes divergences in the model’s noise-predictive distribution. Our method preserves denoising behavior without requiring replay buffers or generative replay. We introduce two variants of our regularization term based on ℓ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> and ℓ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> penalties and integrate them into the unconditional training process. We compare forecasting results on three real-world tasks and evaluate our method against naive sequential training and dropout regularization to demonstrate the effectiveness of functional regularization in mitigating catastrophic forgetting.

  PublisherDOI

• Higher-Order Feature Attribution: Bridging Statistics, Explainable AI, and Topological Signal Processing

  2026-04-21

  articleOpen accessSenior author

  Feature attributions are post-training analysis methods that assess how various input features of a machine learning model contribute to an output prediction. Their interpretation is straightforward when features act independently, but it becomes less clear when the predictive model involves interactions, such as multiplicative relationships or joint feature contributions. In this work, we propose a general theory of higher-order feature attribution, which we develop on the foundation of Integrated Gradients (IG). This work extends existing frameworks in the literature on explainable AI. When using IG as the method of feature attribution, we discover natural connections to statistics and topological signal processing. We provide several theoretical results that establish the theory, and we validate our theory on a few examples.

  PublisherOA PDFDOI

• Assessing Model Proficiency in Autonomous Agents: A Signal Processing Perspective

  2025-04-06

  articleSenior author

  Autonomous agents play a crucial role in applications such as emergency response and urban security, where they are often required to operate independently in critical situations without direct human supervision. A key aspect of autonomy is the agent’s ability to assess its proficiency in carrying out tasks and making decisions based on this assessment. This paper introduces a state-space model to present a novel framework for assessing the proficiency of autonomous agents. The proposed metric is based on statistical model assessment, enabling agents to evaluate their model performance in real-time. More specifically, we focus on the proficiency of the model used for measurement predictions. We validate the effectiveness of our metric through simulations with synthetic data. Future work will explore the potential of this framework to enhance decision-making accuracy and improve task performance.

  PublisherDOI

• Trustworthy Prediction with Gaussian Process Knowledge Scores

  2025-09-08

  articleOpen accessSenior author

  Probabilistic models are often used to make predictions in regions of the data space where no observations are available, but it is not always clear whether such predictions are well-informed by previously seen data. In this paper, we propose a knowledge score for predictions from Gaussian process regression (GPR) models that quantifies the extent to which observing data have reduced our uncertainty about a prediction. The knowledge score is interpretable and naturally bounded between 0 and 1. We demonstrate in several experiments that the knowledge score can anticipate when predictions from a GPR model are accurate, and that this anticipation improves performance in tasks such as anomaly detection, extrapolation, and missing data imputation. Source code for this project is available online at https://github.com/KurtButler/GP-knowledge.

  PublisherOA PDFDOI

• Scalable Random Feature Latent Variable Models

  IEEE Transactions on Pattern Analysis and Machine Intelligence · 2025-07-16 · 1 citations

  articleSenior author

  Random feature latent variable models (RFLVMs) are state-of-the-art tools for uncovering structure in high-dimensional, non-Gaussian data. However, their reliance on Monte Carlo sampling significantly limits scalability, posing challenges for large-scale applications. To overcome these limitations, we develop a scalable RFLVM framework based on variational Bayesian inference (VBI), a deterministic and optimization-based alternative to sampling methods. Applying VBI to RFLVMs is nontrivial due to two key challenges: (i) the lack of an explicit probability density function (PDF) for Dirichlet process (DP) mixing weights, and (ii) the inefficiency of existing VBI approaches when handling the high-dimensional variational parameters of RFLVMs. To address these issues, we adopt the stick-breaking construction for the DP, which provides an explicit and tractable PDF over mixing weights, and propose a novel inference algorithm, block coordinate descent variational inference (BCD-VI), which partitions variational parameters into blocks and applies tailored solvers to optimize them efficiently. The resulting scalable model, referred to as SRFLVM, supports various likelihoods; we demonstrate its effectiveness under Gaussian and logistic settings. Extensive experiments on diverse benchmark datasets show that SRFLVM achieves superior scalability, computational efficiency, and performance in latent representation learning and missing data imputation, consistently outperforming state-of-the-art latent variable models, including deep generative approaches.

  PublisherDOI

• Multi-branch convolutional neural network using intracranial EEG high frequency oscillation features for predicting post-surgical seizure outcomes

  medRxiv · 2025-10-07

  preprintOpen access

  Abstract Pathological high-frequency oscillations (HFOs 80-600 Hz) in intracranial EEG distinguish epileptogenic cortex. However, it is uncertain whether utilizing HFO measures for surgical planning improve epilepsy surgery seizure outcomes and minimize morbidity. The clinical gold standard for planning an epilepsy surgery involves consensus between epileptologists, radiologists, and neurosurgeons based on multimodality findings, and particularly the location of the seizure onset zone. We asked whether seizure freedom following epilepsy surgery could be accurately predicted using machine learning that uses measures of HFO features relative to the boundaries of a surgical resection or laser ablation. We detected and quantified HFOs from depth EEG contacts during 30-200 minutes of non-rapid eye movement sleep from 78 pre-surgical patients from three institutions. We trained a three-branch convolutional neural network (CNN) using 3 neuroanatomic features and 37 HFO derived features. The first and second CNN branches computed within and between patient differences, respectively, and the third branch contains the resected contacts that also influenced branches 1 and 2. We found that this HFO CNN labeled the seizure free patients with 92% accuracy using 5-fold cross-validation. These results suggest that a resection planned with the clinical gold standard can be prospectively evaluated by a HFO CNN approach to test whether the resection boundaries will achieve a seizure free outcome. Future work will explore utilizing the HFO CNN approach for counterfactual virtual resections constrained by a utility function to minimize morbidity.

  PublisherOA PDFDOI

• Ventriculomegaly without elevated intracranial pressure? Normal pressure hydrocephalus as a disorder of the cerebral windkessel

  Frontiers in Neurology · 2025-05-01 · 1 citations

  articleOpen access

  Objective Normal pressure hydrocephalus (NPH) is characterized by ventriculomegaly without elevations in intracranial pressure (ICP). One way of viewing hydrocephalus is as a disorder of the cerebral windkessel. The cerebral windkessel is the system that dampens the arterial blood pressure (ABP) pulse in the cranium, transmitting this pulse from arteries to veins via the cerebrospinal fluid (CSF) path, bypassing the microvasculature to render capillary flow smooth. When the windkessel is physiologically tuned, windkessel effectiveness ( W ) is given by: W = IE / R , where I represents CSF path inertance (pulse magnitude), E is CSF path elastance, and R is resistance in the CSF path. In NPH, we posit that there is a combination of arteriosclerosis (blunting the CSF pulse in the SAS- lowering I ), and age-related softening of brain tissue (decreasing the elastance of subarachnoid CSF pathways- lowering E ). Methods To model the windkessel, we utilize a tank circuit with parallel inductance and capacitance to simulate the pulsatile flow of blood and CSF as alternating current (AC), and smooth flow as direct current (DC). We model NPH as a disorder of windkessel impairment by decreasing windkessel inertance (reflecting diminished CSF pulsatility in the SAS from arteriosclerosis) and decreasing intracranial elastance (reflecting age-related brain atrophy). We simulate ventriculomegaly and shunting by lowering the resistance of this circuit. Results In simulating NPH using this circuit, we found significant elevations in the amplitude and power of AC in the CSF and capillary paths when inertance and elastance were decreased. Conversely, this pulse power decreased with decreased resistance in the CSF path from ventriculomegaly and shunting. Conclusion Simulations of NPH demonstrated increased amplitude and power of AC in the CSF and capillary paths due to windkessel impairment. We posit that this pulsatility is redistributed from the SAS to the ventricular CSF path, exerting pulsatile stress on the periventricular leg and bladder fibers, which may explain NPH symptomatology. Ventriculomegaly may represent an active adaptation to improve windkessel effectiveness by decreasing CSF path resistance to mitigate decreased CSF path inertance and parenchymal elastance. Shunting provides a low resistance, accessory windkessel to obviate adaptive ventriculomegaly. This has significant implications in understanding this paradoxical condition.

  PublisherOA PDFDOI

• Beam Pattern Optimization for Integrated Sensing and Communication in UAV Applications

  2025-03-30

  articleOpen accessSenior author

  The deployment of unmanned aerial vehicles (UAVs) equipped with uniform rectangular array (URA) offers a promising solution for simultaneous communication with multiple ground base stations (GBSs) and the sensing of ground-based target. However, designing a transmit beamforming system that optimizes communication while maintaining high sensing accuracy is challenging. In this regard, this paper proposes a comprehensive beam pattern optimization framework tailored for integrated sensing and communication in UAV applications, where the beam is synthesized based on the required steering direction, beam width, side lobe suppression, and nulling in targeted areas, providing robust connectivity. This comprehensive framework integrates sensing and communication, improving system performance and connectivity in UAV applications.

  PublisherOA PDFDOI

• Improving Communication Performance of Passive Backscattering Tags Using Collaborative Backscatter

  2025-04-22 · 2 citations

  articleSenior author

  We propose collaborative backscatter techniques for tag-to-tag communication between battery-less RF tags. The low incident backscatter power and the limited processing ability in the passive receive circuits limit the performance of such links. By recruiting 'helper' tags to boost backscatter signals, such links can be substantially strengthened, depending upon the network topology and channel conditions. Two techniques are developed and evaluated on a prototype tag network, demonstrating close-to-optimal performance with low computational overhead.

  PublisherDOI

Recent grants

• CIF: Small: Belief Evolutions in Networks of Bayesian Agents

  NSF · $393k · 2013–2017

• CCF: Medium: Inference with dynamic deep probabilistic models

  NSF · $1.2M · 2022–2027

• Rethinking Electronic Fetal Monitoring to Improve Perinatal Outcomes and Reduce Frequency of Operative Vaginal and Cesarean Deliveries

  NIH · $3.2M · 2019–2025

• CIF: Small: Learning and herding in complex systems

  NSF · $470k · 2010–2013

• CPS: Medium: Collaborative Research: Scalable Intelligent Backscatter-Based RF Sensor Network for Self-Diagnosis of Structures

  NSF · $819k · 2021–2026

Frequent coauthors

• Sergios Theodoridis

  National and Kapodistrian University of Athens

  1531 shared

• Anna Scaglione

  Cornell University

  1527 shared

• W.C. Karl

  Boston University

  1524 shared

• J Treichler

  1524 shared

• Alex Acero

  Apple (Israel)

  1524 shared

• Mari Ostendorf

  1524 shared

• Charles A. Bouman

  Purdue University System

  1424 shared

• Alex C. Kot

  1424 shared

Labs

• Electrical and Computer EngineeringPI

Education

• Ph.D., Electrical Engineering

  University of Belgrade

  1989

• M.S., Electrical Engineering

  University of Belgrade

  1985

• B.S., Electrical Engineering

  University of Belgrade

  1981