Dr. Sumit Raurale

Intervention-Aware Temporal Modeling for Pregnancy Risk Progression

This research addressed a fundamental challenge in longitudinal clinical prediction: how to model disease or risk trajectories when medical interventions alter the natural progression of outcomes. In pregnancy care, clinical decisions such as medication adjustments, additional monitoring, or early referral can significantly influence subsequent risk. Traditional predictive models often treat outcomes as static endpoints without accounting for these feedback effects.

In this project, I developed temporal modeling frameworks that integrate clinical guideline knowledge with data-driven learning approaches. The objective was to move beyond isolated risk classification toward dynamic estimation of gestational risk progression over time.

The methodological contributions included:

Time-to-event and survival-based modeling of gestational outcomes
Integration of structured clinical rules within deep temporal architectures
Explicit modeling of time gaps between visits and delayed test availability
Calibration analysis to ensure reliability of predicted risk trajectories

This work contributes to a broader research direction in intervention-aware AI systems, where predictions and clinical actions are jointly considered within a temporal decision framework. The resulting models emphasize interpretability, reliability, and robustness in real-world clinical workflows.

Longitudinal Maternal and Fetal Risk Modeling

This research program addressed a central challenge in pregnancy monitoring: how to model maternal and fetal risk progression over time when clinical data is irregular, partially observed, and distributed across multiple modalities. During gestation, information accumulates through intermittent ultrasound scans, physiological measurements, and structured clinical records. These signals often arrive with delays, missing values, or incomplete context, making conventional modeling approaches insufficient.

I developed temporal and multimodal learning frameworks to estimate fetal growth trajectories and dynamic pregnancy risk progression. The methodological focus was on designing sequence models capable of capturing long-range temporal dependencies under irregular visit schedules. This included transformer-based architectures and continuous-time learning approaches that explicitly account for uneven sampling intervals and delayed observations.

A key contribution of this work was the development of multimodal fusion strategies integrating ultrasound-derived biometric features, fetal ECG signals, and structured electronic medical records within a unified representation framework. Particular attention was given to robustness under partial modality availability, ensuring stable performance even when one data source was missing.

Core methodological contributions include:

Temporal representation learning under irregular and sparse clinical timelines
Survival and time-to-event modeling for gestational risk estimation
Missing-modality-aware multimodal fusion architectures
Robust cross-site validation under heterogeneous hospital settings

The resulting models were evaluated across multi-site clinical cohorts and translated into regulated maternal health monitoring systems. Beyond application, this work contributes to the broader field of longitudinal and trustworthy AI systems for high-stakes decision environments.

Multimodal Chronic Disease Progression and Digital Twin Modeling

This research addressed the challenge of modeling chronic disease progression using heterogeneous and longitudinal clinical data. In real-world settings, patient records often consist of laboratory reports, physiological measurements, imaging biomarkers, and indirect indicators of pathology collected at irregular intervals. These data streams are frequently incomplete and evolve over time, making robust trajectory modeling non-trivial.

I developed multimodal sequence learning architectures capable of integrating diverse data sources within a unified temporal framework. The objective was to construct patient-specific progression models that could continuously update risk estimates as new evidence became available, forming the basis of a digital twin–style monitoring system. Rather than static classification, the models focused on capturing evolving disease dynamics.

Key methodological contributions include :

Design of early and late fusion strategies for heterogeneous biomedical data
Sequence modeling under sparse and partially observed longitudinal records
Missing-modality-aware representation learning
Clinician-aligned disease staging frameworks to ensure interpretability

This work advanced personalized monitoring approaches by linking temporal machine learning with clinically meaningful disease progression modeling. It also contributed to the broader methodological theme of building reliable AI systems under incomplete and evolving real-world data conditions.

Wearable Health Monitoring and Stress Modeling

This research focused on developing multimodal wearable sensing systems for continuous health monitoring under real-world conditions. Unlike laboratory environments, wearable devices operate in dynamic and uncontrolled settings where physiological signals are influenced by motion, environmental factors, and user variability. Designing reliable models under such conditions requires careful integration of signal processing and machine learning.

In Industry, I worked on physiological modeling frameworks using biosignals such as photoplethysmography (PPG), galvanic skin response (GSR), accelerometry, and skin temperature. A central challenge was distinguishing genuine stress-related physiological responses from motion-induced artifacts and contextual noise.

The methodological focus included :

Multimodal fusion of heterogeneous biomedical data
Noise-robust time-series modeling under motion and environmental variability
Continuous risk scoring from streaming wearable data
Embedded and deployment-aware AI pipeline design

This work strengthened my expertise in real-time physiological signal modeling and robust temporal learning under non-controlled conditions. It also contributed to my broader research direction of designing AI systems that remain reliable outside ideal laboratory settings.

Neonatal Brain Injury Assessment and Seizure Prediction

During my postdoctoral research, I focused on developing AI-driven methods for automated assessment of neonatal brain injury and early seizure prediction in intensive care settings. Neonatal neurocritical care presents a uniquely challenging modeling problem: physiological signals are high-dimensional, noisy, and temporally evolving, while imaging data varies across hospitals and acquisition protocols.

I developed multimodal deep learning frameworks integrating electroencephalography (EEG) and magnetic resonance imaging (MRI) data to support both injury grading and long-term outcome prediction. For seizure prediction, I designed temporal models that combined time–frequency representations of EEG with recurrent and attention-based sequence architectures to capture evolving neural dynamics.

A central challenge in this work was ensuring robustness across heterogeneous clinical environments. To address this, I focused on representation learning strategies capable of generalizing across sites, signal variability, and imaging differences.

Core methodological contributions include :

Multimodal integration of EEG and MRI within unified deep learning frameworks
Temporal sequence modeling for evolving neonatal risk trajectories
Time–frequency feature extraction and structured signal representations
Uncertainty-aware prediction strategies to reduce false alarms in critical care

This work resulted in peer-reviewed publications in Journal of Neural Engineering, Nature: Scientific Data, EMBS EMBC, EUSIPCO and was validated across multi-site European neonatal cohorts. Beyond application, it strengthened my research direction in temporal and trustworthy AI for high-stakes medical decision systems.

Automated Neonatal Brain State Characterization

This research focused on developing automated methods to characterize neonatal brain states using electroencephalography (EEG). In newborns, especially those in intensive care, EEG patterns such as burst–interburst activity and tracé alternant reflect critical information about brain maturation and neurological health. However, manual interpretation is time-consuming and requires specialized expertise.

I developed signal-processing and machine learning frameworks to automatically detect and quantify these clinically significant EEG patterns. The work combined time–frequency spectral analysis with structured segmentation algorithms to distinguish bursts and inter-burst intervals and to identify tracé alternant states associated with quiet sleep and developmental progression.

Core methodological contributions include :

Time–frequency spectral feature modeling for neonatal EEG
Automated burst and inter-burst segmentation algorithms
Developmental brain state classification frameworks
Extraction of clinically interpretable EEG biomarkers

This work contributed to advancing automated tools for neonatal neurodevelopmental assessment and strengthened my expertise in physiological signal modeling, interpretable machine learning, and clinically grounded AI systems.

Real-Time Physiological AI Systems for Wearable Platforms

My doctoral research centered on designing real-time physiological AI systems for wearable human–computer interaction and secure authentication. The core scientific challenge was to develop machine learning models that remain reliable outside controlled laboratory environments, where sensor placement, hardware configuration, and environmental conditions vary significantly.

Using electromyography (EMG) as the primary signal source, I investigated how to build sensor-agnostic modeling frameworks capable of accurate motion recognition even when electrodes were placed randomly on the forearm. This required integrating robust signal processing with time-series learning architectures that could tolerate device variability, noise, and user-specific differences.

Beyond algorithm development, a major focus of my PhD was system-level deployment. I designed end-to-end pipelines spanning data acquisition, feature extraction, model training, and embedded inference, enabling real-time operation on low-power wearable hardware. This work bridged theoretical modeling with practical implementation constraints.

The thesis contributed to several methodological themes :

Sensor-agnostic EMG representation learning
Robust time-series modeling under hardware and environmental variability
Embedded real-time inference architectures for wearable systems
Physiological biometric modeling for secure authentication

This research was disseminated through peer-reviewed publications, including IEEE Transactions on Signal Processing, IEEE ICASSP, IEEE Access and validated in real-time embedded environments. It established my long-standing interest in reliable AI systems that function under real-world constraints, a theme that continues to shape my work in longitudinal and trustworthy AI.

Rehabilitation Monitoring and Biomechanical Modeling

My early research investigated longitudinal monitoring of post-surgical rehabilitation using physiological and biomechanical signals. Focusing on patients recovering from knee replacement surgery, I developed machine learning models to track recovery trajectories using electromyography (EMG) and movement-derived features.

A central objective was to move beyond subjective recovery assessment toward quantitative, clinician-aligned progression modeling. This required structured physiological data acquisition protocols and statistical learning frameworks capable of capturing temporal changes in muscle activation and motor performance.

Key methodological themes included :

Temporal modeling of rehabilitation trajectories
Quantitative recovery assessment aligned with clinical metrics
Signal processing and feature extraction from EMG and movement data
Structured experimental design for physiological data collection

This work established my long-term research interest in modeling health progression over time and provided the conceptual foundation for my later contributions in longitudinal, multimodal, and uncertainty-aware AI systems.