Publications

Artificial Intelligence (AI) and digital devices are becoming more pervasive in the monitoring of individuals’ health as device functionalities increase and their presence in daily life continues to grow. Subsequently, these tools can provide transformative healthcare practice by enabling more personalized, proactive, and engaging interventions. This work presents the design and development of an interactive physical therapy system for adolescents that integrates a wearable EMG leg sensor array with a computer-based video game. The system captures EMG signals during therapeutic exercises and uses them to drive in-game mechanics, providing real-time feedback while also aiming to improve therapeutic engagement. A key feature of the system is its ability to adapt based on individual performance and clinician input, allowing personalized therapy progression and reporting. Using this approach, the system supports continuous personalized assessment and customized intervention for children while promoting motivation through dynamic gameplay. The developed system illustrates a promising direction for designing interactive, AI-augmented health technologies that can enhance pediatric rehabilitation outcomes through accessible, game-based experiences. Further, this work contributes to ongoing research in human-computer interaction, AI, and behavioral health by exploring multisensory interactions, social AI elements, and methods for evaluating user engagement.

Traumatic brain injury (TBI), particularly mild forms resulting from blast exposure, remains a diagnostic challenge among veterans due to delayed symptom onset and overlapping psychological conditions such as posttraumatic stress disorder. Current diagnostic methods rely heavily on subjective assessments, contributing to underdiagnosis and inconsistent care. Electroencephalography (EEG) offers a non-invasive, real-time measure of cortical activity with high temporal resolution, making it a promising tool for objective TBI assessment. This sub-study, nested within a randomized controlled trial evaluating hyperbaric oxygen therapy (HBOT) for veterans with TBI, investigates EEG as both a diagnostic modality and a therapeutic monitor. This sub-study adopts the same triple-blinded, randomized, parallel-group design as the parent trial. Participants are veterans with mild to moderate TBI, randomized to receive either HBOT or a sham treatment. EEG data are collected at three time points: baseline (pre-treatment), midpoint (18–21 dives), and post-treatment (2 weeks after completion). EEG recordings are performed during standardized cognitive and motor tasks using a 32-channel wireless headset. Primary outcomes include event-related potential amplitudes and alpha-band spectral power, analyzed for longitudinal changes and group differences. Secondary outcomes include latency measures, spectral power across additional frequency bands, and functional connectivity metrics. Data are modeled using repeated measures analysis to assess treatment effects and individual trajectories. This sub-study aims to validate EEG as a scalable and objective tool for diagnosing and monitoring TBI in clinical settings. By identifying electrophysiological signatures associated with injury severity and treatment response, EEG may enhance diagnostic precision and support personalized care strategies. The integration of EEG within a larger therapeutic trial framework allows for comprehensive evaluation of its clinical utility. Findings may inform future protocols for TBI assessment and contribute to the development of neurophysiological biomarkers that complement existing symptom-based approaches.

One of the most promising avenues of artificial intelligence (AI) integration into medicine is its examination, evaluation, and characterization of pathological slides. The use of large language models (LLMs), the AI model subtype that is becoming increasingly popular, in pathological applications remains unexplored. This study investigates the histological image recognition capabilities of the multimodal models Gemini 1.5 Flash, ChatGPT-4o, and Claude 3.5 Sonnet and assesses their suitability for clinical or medical education use. The models were evaluated using 300 digital histology images derived from the University of South Florida Morsani College of Medicine Virtual Microscopy database, with a prompt to ascertain each model's ability to identify tissue type and plane of sectioning used. The images included the three subtypes in both longitudinal and transverse planes of sectioning. Standard machine learning metrics such as precision, recall, accuracy, and F1 score were used to classify and evaluate each model’s abilities. In the prediction of tissue type, OpenAI’s ChatGPT had the highest metrics with an F1 score of 0.772, while Claude yielded an F1 score of 0.380, and Gemini produced a 0.460 F1 score. In the prediction of sectioning, ChatGPT produced an F1 score of 0.396, while Claude produced a value of 0.472, and Gemini yielded 0.344. Overall, the results indicate that ChatGPT is most effective at identifying tissues. However, the inaccuracy demonstrated in evaluating sectioning compared to other models leaves room for improvement in its overall accuracy across varying tissue samples to reliably supplement medical education or clinical use.

Anterior cruciate ligament (ACL) rupture is a critical concern in sports medicine. This study presents a detailed evaluation of machine learning (ML) techniques for the prediction of anterior cruciate ligament (ACL) injuries - a critical concern in sports medicine that often entails prolonged recovery periods and significant patient burden. Leveraging the transformative potential of artificial intelligence (AI) in medical applications, our work assesses eight distinct ML models (i.e., Support Vector Machines, Decision Tree Classifiers, Random Forest, Stochastic Gradient Descent, Logistic Regression, Gradient Boosting, Ridge Regression, and Linear Discriminant Analysis). Models were trained and tested on four datasets: 53-feature (Binary) ACL Rupture Biomechanical and Demographic (ARBD/BARBD) and 13-feature (Binary) ACL Rupture Wearable (ARW/BARW) that are readily accessible in real-life scenarios. Models were evaluated under a three-class schema ’pre-rupture,’ ’trial prior to rupture,’ ’rupture’ and then reclassified into a binary ’pre-rupture’ vs. ’elevated risk.’ Our analyses reveal that early-phase force metrics, particularly those recorded at 33 milliseconds (e.g., 33ms_Fx and 33ms_Fz) with ARBD and BARBD datasets and initial-contact forces (e.g., IC_Fx, IC_Fz) over demographic variables in ARW and BARW datasets, consistently emerge as significant predictors of injury across multiple and binary models. Notably, across the ARBD and ARW datasets, the ML models achieved accuracies ranging from approximately 79% to 87%, which improved markedly to a range of 92% to 95% when reclassified into a binary classification. These findings underscore the clinical relevance of early dynamic measurements and demonstrate the robustness and generalizability of our approach.

Speech emotion recognition (SER) is an important application in Affective Computing and Artificial Intelligence. Recently, there has been a significant interest in Deep Neural Networks using speech spectrograms. As the two-dimensional representation of the spectrogram includes more speech characteristics, research interest in convolution neural networks (CNNs) or advanced image recognition models is leveraged to learn deep patterns in a spectrogram to effectively perform SER. Accordingly, in this study, we propose a novel SER model based on the learning of the utterance-level spectrogram. First, we use the Spatial Pyramid Pooling (SPP) strategy to remove the size constraint associated with the CNN-based image recognition task. Then, the SPP layer is deployed to extract both the global-level prominent feature vector and multi-local-level feature vector, followed by an attention model to weigh the feature vectors. Finally, we apply the ArcFace layer, typically used for face recognition, to the SER task, thereby obtaining improved SER performance. Our model achieved an unweighted accuracy of 67.9% on IEMOCAP and 77.6% on EMODB datasets.

Due to the complicated nature of Parkinson disease (PD), a number of subjective considerations (eg, staging schemes, clinical assessment tools, or questionnaires) on how best to assess clinical deficits and monitor clinical progression have been published; however, none of these considerations include a comprehensive, objective assessment of all functional areas of neurocognition affected by PD (eg, motor, memory, speech, language, executive function, autonomic function, sensory function, behavior, and sleep). This paper highlights the increasing use of digital health technology (eg, smartphones, tablets, and wearable devices) for the classification, staging, and monitoring of PD. Furthermore, this Viewpoint proposes a foundation for a new staging schema that builds from multiple clinically implemented scales (eg, Hoehn and Yahr Scale and Berg Balance Scale) for ease and homogeneity, while also implementing digital health technology to expand current staging protocols. This proposed staging system foundation aims to provide an objective, symptom-specific assessment of all functional areas of neurocognition via inherent device capabilities (eg, device sensors and human-device interactions). As individuals with PD may manifest different symptoms at different times across the spectrum of neurocognition, the modernization of assessments that include objective, symptom-specific monitoring is imperative for providing personalized medicine and maintaining individual quality of life.

Speech signal analysis to support objective clinical decision-making has gained immense interest, especially in neurological disorders. This research assessed the feasibility of speech analysis on the detection of concussions. Using a speech dataset from 82 concussed and 82 healthy participants, we extracted two speech feature sets focusing on Mel Frequency Cepstral Coefficients (MFCCs) to characterize speech articulation. A machine learning pipeline was developed to discriminate concussion speech from healthy speech by applying Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Decision Tree (DT) classifiers. All three classifiers trained on the MFCC-based feature set achieved Matthew's correlation coefficient score above 0.5 on the holdout data set. DT model achieved a 78% sensitivity and 75% specificity. The findings of this research serve as proof-of-concept for speech analysis of concussion detection.

BACKGROUND: Digital biomarkers are increasingly used in clinical decision support for various health conditions. Speech features as digital biomarkers can offer insights into underlying physiological processes due to the complexity of speech production. This process involves respiration, phonation, articulation, and resonance, all of which rely on specific motor systems for the preparation and execution of speech. Deficits in any of these systems can cause changes in speech signal patterns. Increasing efforts are being made to develop speech-based clinical decision support systems.

OBJECTIVE: This systematic scoping review investigated the technological revolution and recent digital clinical speech signal analysis trends to understand the key concepts and research processes from clinical and technical perspectives.

METHODS: A systematic scoping review was undertaken in 6 databases guided by a set of research questions. Articles that focused on speech signal analysis for clinical decision-making were identified, and the included studies were analyzed quantitatively. A narrower scope of studies investigating neurological diseases were analyzed using qualitative content analysis.

RESULTS: A total of 389 articles met the initial eligibility criteria, of which 72 (18.5%) that focused on neurological diseases were included in the qualitative analysis. In the included studies, Parkinson disease, Alzheimer disease, and cognitive disorders were the most frequently investigated conditions. The literature explored the potential of speech feature analysis in diagnosis, differentiating between, assessing the severity and monitoring the treatment of neurological conditions. The common speech tasks used were sustained phonations, diadochokinetic tasks, reading tasks, activity-based tasks, picture descriptions, and prompted speech tasks. From these tasks, conventional speech features (such as fundamental frequency, jitter, and shimmer), advanced digital signal processing-based speech features (such as wavelet transformation-based features), and spectrograms in the form of audio images were analyzed. Traditional machine learning and deep learning approaches were used to build predictive models, whereas statistical analysis assessed variable relationships and reliability of speech features. Model evaluations primarily focused on analytical validations. A significant research gap was identified: the need for a structured research process to guide studies toward potential technological intervention in clinical settings. To address this, a research framework was proposed that adapts a design science research methodology to guide research studies systematically.

CONCLUSIONS: The findings highlight how data science techniques can enhance speech signal analysis to support clinical decision-making. By combining knowledge from clinical practice, speech science, and data science within a structured research framework, future research may achieve greater clinical relevance.

OBJECTIVES:  This research study aims to advance the staging of Parkinson's disease (PD) by incorporating machine learning to assess and include a broader multifunctional spectrum of neurocognitive symptoms in the staging schemes beyond motor-centric assessments. Specifically, we provide a novel framework to modernize and personalize PD staging more objectively by proposing a hybrid feature scoring approach.

METHODS:  We recruited 37 individuals diagnosed with PD, each of whom completed a series of tablet-based neurocognitive tests assessing motor, memory, speech, executive functions, and tasks ranging in complexity from single to multifunctional. Then, the collected data were used to develop a hybrid feature scoring system to calculate a weighted vector for each function. We evaluated the current PD staging schemes and developed a new approach based on the features selected and extracted using random forest and principal component analysis.

RESULTS:  Our findings indicate a substantial bias in current PD staging systems toward fine motor skills, that is, other neurological functions (memory, speech, executive function, etc.) do not map into current PD stages as well as fine motor skills do. The results demonstrate that a more accurate and personalized assessment of PD severity could be achieved by including a more exhaustive range of neurocognitive functions in the staging systems either by involving multiple functions in a unified staging score or by designing a function-specific staging system.

CONCLUSION:  The proposed hybrid feature score approach provides a comprehensive understanding of PD by highlighting the need for a staging system that covers various neurocognitive functions. This approach could potentially lead to more effective, objective, and personalized treatment strategies. Further, this proposed methodology could be adapted to other neurodegenerative conditions such as Alzheimer's disease or amyotrophic lateral sclerosis.