1,721,025 research outputs found
Recommended from our members
Prediction, Interpretation and Counterfactual Generation for Acute Care Applications
No full textOrgan failure in critically ill patients is a dynamic process which evolves over time as physiological stateschange and multiple organ systems interact. Anticipating these changes and understanding what drives them is essential for improving patient outcomes, yet current tools often lack the ability to provide both accurate forecasts and clinically meaningful insights. There is a growing need to move beyond prediction to explore “what if” scenarios that can guide clinicians to treatment decisions and offer actionable strategies at the bedside. This work presents an integrated framework for advancing AI-related translational sciences that addresses these gaps through three components. First, we developed predictive models using longitudinal clinical data to forecast the onset and progression of organ dysfunction and offer early warnings of deterioration. Second, we applied interpretable methods to identify the clinical variables and temporal patterns that are most influential to model outputs, ensuring transparency and alignment with biomedical domain knowledge. Third, we generate counterfactual patient trajectories to explore how changes in treatments or physiological states may have influenced outcomes. Applied on real-world intensive care datasets, the framework demonstrate strong predictive performance, interpretable explanations that clinicians find meaningful, and counterfactual scenarios that are both plausible and actionable. Together, these contributions connect model development to prospective clinical evaluation, creating a unified approach that integrates prediction, interpretation, and counterfactual reasoning to enable more informed and timely decision making in critical care.Dissertation not available (per author’s request
Recommended from our members
Automated Analysis of Scattering-based Light Sheet Microscopy Images of Anal Squamous Intraepithelial Lesions
No full textAnal cancer presents diagnostic challenges, particularly in identifying high-grade squamous intraepithelial lesions (HSIL), with its increasing incidence and mortality rates. Current diagnosis methods, including cytology, biopsy, and high resolution anoscopy (HRA), provide important diagnostic information. However, cytology is often limited by suboptimal sensitivity and specificity, while high resolution anoscopy-guided biopsy is limited by its long processing times due to unnecessary biopsies and staining requirements. Scattering-based light sheet microscopy (sLSM) can offer an alternative approach by utilizing intrinsic tissue scattering properties to visualize morphologic features without the need for additional labeling or staining.In this study, we developed and evaluated an automated algorithm for analyzing 187 sLSM images obtained from 80 anal biopsies. The method employed a row-by-row binarization technique for nuclear segmentation, achieving high precision (0.97) and recall (0.91). Seven nuclear features, including nuclear intensity, intensity slope as a function of depth, nuclear-to-nuclear distance, nuclear-to-cytoplasm ratio, cell density, nuclear area, and proportion of pixels corresponding to nuclei were extracted and statistically analyzed. Among the seven features, six showed statistically significant differences between HSIL and non-HSIL (non-dysplastic or low-grade squamous intraepithelial lesion, LSIL). A linear support vector machine (SVM) was trained and tested using five-fold cross validation on these features. The classifier achieved a sensitivity of 90%, specificity of 70%, and area under the curve (AUC) of 0.89 for per-image diagnosis, and sensitivity of 90%, specificity of 80%, and area under the curve (AUC) of 0.93 for per-biopsy diagnosis
Recommended from our members
Informatics for Trauma-Mediated Psychiatric Illness
No full textTraumatic events such as direct or indirect exposure to serious injury or situations inciting strong fear response have been shown to result in mental disorders that are complex to diagnose, prognosticate, and treat. Neuropsychiatric sequelae following head injuries, including post concussive syndrome (PCS) and post-traumatic stress disorder (PTSD) are particularly challenging because of overlapping symptoms and the profound nature of the injury itself. This work aims to elucidate the similarities and differences between PCS and PTSD, two common sequelae following traumatic brain injury, using an emerging framework (Research Domain Criteria – RDoC) for understanding mental health disorders, and better characterize the mechanisms leading to psychiatric illness as a result of neurotrauma. The application of RDoC is demonstrated using analyses of clinical data from the Alzheimer's Disease Neuroinitiative. This analysis involves identification of subtle symptom variants for differentiating sequelae when clinical prognosis is uncertain based self-reported symptoms. In addition to specific self-reported symptoms, features in positron emission tomography and diffusion tensor imaging are also identified. These results are further supplemented by identification of brain structures associated with self-reported symptoms using imaging data. Overall, this project demonstrates the application of RDoC to better characterize PCS and PTSD, which would potentially allow for more effective treatment and management of these disorders.Release after 08/22/202
Recommended from our members
Application of Machine Learning Techniques for Prognosis of Traumatic Brain Injury Patients in Intensive Care Units
No full textWith advances in digital health technologies and proliferation of big biomedical data in recent years, applications of Machine Learning (ML) in healthcare and medicine have gained significant attention. Modern Intensive Care Units (ICUs), in particular, are equipped to generate rich multimodal clinical data on critically-ill patients. In this thesis, we focus on applying machine learning techniques for prognostication of Traumatic Brain Injury (TBI) patients in ICU, which is the leading cause of death and disability among children and adults of age less than 44. We present two case studies to demonstrate the feasibility and applicability of machine learning techniques: one for mortality prediction in TBI patients and the second for extracting patterns from physiological data collected from TBI patients. For the case study I, clinical data including demographics, vital signs, and physiological data for the first 72 hours of TBI patients were extracted from the Medical Information Mart for Intensive Care III (MIMIC III) database. Several traditional supervised machine learning algorithms such as artificial neural network, support vector machine, and logistic regression were employed to construct prediction models. Bagging and Voting techniques were implemented to improve the performance of these algorithms. By comparing the performances of these algorithms, we showed that deploying voting techniques on several different ML models can improve the overall performance. These algorithms obtained the highest Area Under receiver operating characteristic Curve (AUC) of 0.91. For the case study II, an exploratory, secondary analysis of physiologic data of TBI patients from the Phase III trial of Progesterone for Traumatic Brain Injury, Experimental Clinical Treatment (PROTECT) was performed. Subspace clustering was used to extract relationships between various physiologic variables. For both studies, 10-fold cross validation was used for evaluation purposes.Release after 24-May-202
Recommended from our members
New Platform Using eLAMP 30 Degree Mie Scatter for Affordable and Efficient miR-21 Detection and Quantification
No full textMicroRNAs are an upcoming biomarker for screening several different diseases. MicroRNAs (miRs) are single-stranded RNA molecules of 15-27 nucleotide lengths that control gene expression during the translational level. miR-21 specifically is one of the earliest cancer detectors, targeting numerous suppressor genes associated with proliferation, apoptosis, and invasion [1]. Loop-mediated isothermal amplification (LAMP) is an isothermal nucleic acid amplification method that is rapid, sensitive, and efficient at the identification of diseases [2]. Introducing an emulsion environment within the LAMP assay creates pockets for the reaction, which during amplification, will cause a decrease in droplet size and a change in light scatter. To identify this phenomenon, it was found that a mini spectrometer and a hot plate could detect changes in the eLAMP reactions that can be quantified for a target molecule at 30 degrees. The spectrometer and the hotplate can cost upwards of 5000-6000 dollars to acquire and measure intensity changes that are not always accurate and require recalibration. There is a need for an extremely affordable and more accurate way to measure Mie scatter eLAMP, so that the field can be researched further and become more readily available for in-home testing. The light intensity can be easily measured through inexpensive materials with proper orientation and light suppression. Through a new design utilizing 30-degree Mie scatters and eLAMP, the presence of mir-21 can be detected with an Arduino, photoresistor, and LED. The heating element and vortex flow can also be created for much less than the cost of a hotplate using a PID temperature controller, fan, and Peltier Heater. Altogether, the device can run a MIE scatter eLAMP reaction and detect and quantify the presence of mir-21 for approximately 100 dollars
Recommended from our members
Sustainable Biorecovery of Critical Materials From Lithium-Ion Batteries: Techno-Economic Analysis, Life Cycle Assessment, Supply Chain Optimization
No full textThe demand for lithium-ion batteries (LIBs) has surged in recent years, owing to their excellent electrochemical performance and increasing adoption in electric vehicles and renewable energy storage. As a result, the expectation is that the primary supply of LIB materials (e.g., lithium, cobalt, and nickel) will be insufficient to satisfy the demand in the next five years, creating a significant supply risk. The number of LIBs reaching their end-of-life (EOL) is expected to grow substantially in the next decade. These EOL LIBs represent a significant secondary source of materials that can be recovered and reused in LIBs or other products.Bioleaching has received substantial attention in recent years for its potential to recover metals in a more environmentally sustainable manner than conventional hydrometallurgical and or pyrometallurgical methods. Developing a cost-effective LIB bioleaching process could be a promising alternative to traditional energy-intensive technologies. The purpose of this study is to increase the bioleaching technology’s readiness for industrial adoption to recycle and recover value from spent LIBs in the United States by evaluating economic and environmental feasibility of the process, optimizing the process, and identifying the optimal supply chain configuration for the LIB bioleaching.
First, techno-economic analysis (TEA) and life cycle assessment (LCA) evaluated economic viability and environmental sustainability of a novel bioleaching technology for value recovery from EOL LIB black mass, i.e., cathode-containing powder, under industrially relevant conditions. Black mass was leached using a biolixiviant produced from corn stover by Gluconobacter oxydans bacteria. Iron(II) was used as a reducing agent to promote metal dissolution. TEA estimated a potential average profit margin of 21% for processing 10,000 tons of black mass per year, which represents approximately 30% of the available black mass in the US in 2020. LCA demonstrated that bioleaching of spent LIBs could be more environmentally sustainable than alternative hydrometallurgical recovery methods such as hydrochloric acid leaching (16−19 kg vs. 43−91 kg CO2 equivalent global warming potential per kg of recovered cobalt). The TEA results are highly dependent on the cost of black mass production, which varies by EOL LIB collection and transportation costs. Emerging technologies for deactivating used LIBs for fire safety at collection centers will allow the transport of EOL LIBs as non-hazardous materials, lower the cost of preparing black mass and thereby increase economic prospects for EOL LIBs recycling using this approach.
Second, another study described the process optimization of the bioleaching conditions for maximum economic competitiveness through design of experiments using iterative response surface methodology (RSM). After two iterations of RSM, i.e., (1) fractional factorial design and steepest ascent (2) central composite design and ridge analysis, the optimal condition for leaching black mass was identified as 2.5% pulp density in 75 mM gluconic acid biolixiviant at 55°C for 30 h. This condition recovered 57% to 84% of the nickel, 71% to 86% of the cobalt, 100% of the lithium, and 100% of the manganese, yielding an estimated positive net profit margin of 17%–26%.
Third study evaluated the potential of bioleaching technology as a sustainable solution for recycling spent LIBs to help inform decision-making processes for stakeholders involved in LIB recycling supply chains. A supply chain model has been developed to include required upstream processes with the objective of maximize economic feasibility of LIB recycling through the technology. The model has been applied for the United States and optimal supply chain configuration was identified, considering the main affecting factors on the economic viability of the technology. The net present value of the supply chain is estimated to be 2.6/kg without a reduction in processing capacity, but prices higher than $3.6/kg significantly decrease the processing capacity of the supply chain, jeopardizing its economic sustainability. The study also examined the non-cooperative scenarios where each tier tries to maximize its own profit to demonstrate how the overall profitability of the supply chain changes with different pricing strategies of sortation facilities and acid producers. The study estimated that the maximum increase in non-recyclable paper and acid prices that the supply chain can tolerate are 400% and 250%, respectively, beyond which the supply chain is no longer sustainable
Recommended from our members
Next-Generation Computational Phenotyping with Large Language Models
No full textIn this dissertation, we presented a critical re-examination of computational phenotyping, a foundational activity in biomedical informatics that supports cohort discovery, observational research, and clinical quality improvement. Despite the development of numerous computable phenotypes across a wide range of clinical outcomes and conditions, the field continues to rely on labor-intensive methods involving manual review and algorithm design. In response, we introduced novel phenotyping methods using Large Language Models (LLMs) to reduce human burden and achieve synergy between human expertise and machine intelligence. These methodological enhancements enabled successful application of LLMs to phenotyping processes previously requiring substantial human oversight. Our work lays the groundwork for the next-generation of computational phenotyping methods, redefining how clinical knowledge is extracted and applied in the era of artificial intelligence. Each of the studies presented in this dissertation supported the progression of next-generation phenotyping methods by assessing the application of LLMs to computational phenotyping tasks. In the first study, we presented PHEONA (Evaluation of PHEnotyping for Observational Health Data), an evaluation framework specifically for LLMs. The components of this framework allowed us to thoroughly evaluate the suitability and feasibility of LLMs for various computational phenotyping tasks. In the second study, we developed a companion framework, SHREC (SHifting to language model-based REal-world Computational phenotyping), that outlined both an end-to-end phenotyping pipeline and the steps necessary to advance next-generation phenotyping methods. Using this framework, we assessed LLMs for concept classification and phenotyping of encounters, which were both individual steps within the end-to-end pipeline. Finally, in the third study, to further evaluate performance deficiencies in applying LLMs to these tasks, we enhanced PHEONA to include an assessment of faulty reasoning within LLM responses.Release after 09/05/202
Recommended from our members
Automated Knowledge Discovery in Healthcare from Complex Data with Covariates
No full textIn this dissertation, the Phase-Type (PH) distribution is studied with consideration of covariates to model and investigate the patient flow information. The length-of-stay (LOS) data of patients with distinct diseases are analyzed in terms of the complete hospital stay or the visit in each department respectively. By PH distribution modeling, the patients clustering and the influences of each covariate are obtained. Comparisons and evaluations among different diseases, different LOS groups, distinct transfer routes and impacts of covariates are implemented.
First, we apply the Coxian PH distributions for fitting the patient flow information collected in a hospital of patients of distinct diseases, including headache, liveborn infant, alcohol abuse, acute upper respiratory infection, and secondary cataract. Based on the results obtained by fitting Coxian PH distributions to the LOS data, the patients can be divided into different groups. The sharing common characteristics, readmission rates, and discharge destinations are evaluated and compared among different disease.
Second, we use the Coxian PH distributions with covariates to fit the patient flow information of both geriatric patients and Alcohol use disorder (AUD) patients collected in a hospital. The influences of the covariates of age, gender, admission type, admit source, and financial class on LOS are assessed and compared through Expectation-Maximization (EM) algorithms. The estimated distributions can classify those patients into different LOS groups and the estimated coefficients and the statistical significance of covariate effects are also achieved and analyzed.
Last, the Coxian PH distributions are integrated in an aggregated Markov chain to model the sequences of LOS in each department that geriatric patients visit during their hospital stay. The aggregated PH distribution is fitted to the intra-hospital transfer flow route by using Maximum Likelihood Estimation (MLE) to obtain the transition rates within each department and transfer rates among departments. The associations between shared characteristics and the transfer routes are also verified.
We conclude that the analysis of associations between LOS groups and readmission rates can help avoid waste of sources. The top predictors are also studied with respect to practical reasons, which may provide guideline for decision-making and resource allocation in healthcare field. The intra-transfer routes within a hospital are associated with the characteristics of patients at admission and discharge.Release after 05/29/202
Recommended from our members
Dynamic Data-Driven Simulation-Based Decision Support System for Medical Procedures
No full textMedical procedures are often performed under the most challenging work environments, demanding higher accuracy and precision in cognitive and technical (psychomotor) skills. More specifically, minimally invasive procedures require higher dexterity to perform the treatment/diagnosis using camera equipment in one hand, and surgical/procedural equipment in the other hand. Thus, it poses a grand challenge to provide a comprehensive system that can provide progressive training, perform objective assessment of psychomotor skills, and assist in studying the decision-making under time and situational uncertainties. To overcome these challenges, a Dynamic Data-Driven Simulation-based decision support system is designed and developed for medical procedures in this dissertation research. To effectively cover the wide spectrum of medical procedures, two types of medical procedures are considered: surgical procedure (Functional Endoscopic Sinus Surgery (FESS)) and emergency procedure (Airway Management). Four particular objectives of this dissertation are to: (1) design and develop a high-fidelity VR-based simulation environment for procedural training and planning, (2) ensure the system calibration from two aspects including positional and parameteric for accurate tracking and realistic representation of anatomical structures, (3) devise computationally efficient data-driven simulation to handle soft-tissue deformation during execution, (4) perform objective proficiency assessment utilizing high-fidelity sensory data acquired during trainee’s performance, (5) conduct the thorough literature review to understand decision-making behavior of caregivers under time and situational uncertainties, and (6) develop a formal process representation while identifying key decision points, and examine a the decision-theoretic model to study the decision-making of caregivers. Firstly, to develop a foundation and medium of feedback delivery to the caregivers, a systems engineering framework (V-model) was utilized in designing a VR-based simulation with a high-fidelity operating room environment with all critical features, including procedural tools, patients, and relevant equipment. In specific, real data of patient-specific anatomical structures (CT scans) and commercially used surgical/procedural tools have been utilized at a system design phase to provide patient-specific and procedure-specific simulations to the trainees. For the system implementation phase, CAD modelling and CT segmentation have been used to develop 3D models of procedure-specific tools and anatomical structures, respectively. The CAD models have also been imported into the Physics-based simulation platform (Unity 3D), where the virtual models have been overlaid over the physical models (3D-printed). Secondly, system calibration for positional and parameteric aspects were achieved to facilitate realistic feel to the user regarding equipment handling and haptic feedback based on the physical interaction between the procedure-specific tools and anatomical structures. Thirdly, a data-driven approach based on sensory data was devised to achieve a computationally efficient method to handle and realistically represent the soft-tissue deformation for the procedure. Fourthly, a hierarchical task analysis of a medical procedure was developed for data acquisition of relevant sensory data (HTC Vive controllers) at the procedural, task, and surgeme levels to provide a robust objective assessment framework. A comprehensive framework was developed for performance assessment based on collected discrete and continuous data for a medical procedure. In specific, dynamic time warping (DTW) algorithm was implemented for data-preprocessing before the actual classification. Hence, DTW was utilized in two folds: (1) to obtain the estimated motion trajectory of an expert (due to lack of ground truth pertaining to ideal procedural motion), (2) to acquire the distance measures pertaining to all attributes of the trainee’s motion by comparison with (1). The output distance measures corresponding to each attribute served as an input to the classification algorithm (decision tree) to classify the trainees based on their current proficiency level (novice and experts). Finally, comprehensive literature review was conducted to gain deeper insights on decision-making behavior of caregivers during a medical procedure, which helped in achieving a formal process modeling of a medical procedure was performed to facilitate research studies understanding the cognitive behavior of caregivers. Specifically, process planning and execution representation using AND/OR graphs was derived for a medical procedure to capture key decision/task alternatives and sequencing problems within the caregiver’s decision-making. Moreover, discussions on the caregiver's decision-making strategy and the application of the decision-theoretic model (R-DFT) for solving AND/OR junctions based on risk and time-urgency have been proposed. The utilization of R-DFT allows the incorporation of intuitionistic and deliberative decision making under a unified framework. The works developed in this dissertation research have been implemented and validated with numerous experiments comprising expert and novice users for two applications (i.e., FESS and Airway Management). Dissemination of the developed works in resident training and study protocols will significantly impact a steady progression of trainee’s skills and gain deeper insights behind decision-making behavior of caregivers. Furthermore, the developed work will open avenues in digital education by offering cost-effective methods for medical education.Release after 08/20/202
- …
