Graduate

Precision Care Medicine: EN.580.680-681

This two-semester design course incorporates the paradigm of modeling, personalization, and application. Precision Care Medicine is a combined project and lecture course, with projects proposed by faculty of the Anesthesiology and Critical Care Medicine (ACCM) Department and their collaborators.

Students assemble into teams and collaborate with their project mentors to develop a work plan and apply methods of computational medicine to solve important health care problems. Teams are also charged with designing, validating, and deploying a web-application that delivers the computational method for solving the underlying healthcare problem to the user. HIPAA regulations, use of human subject data, and requirements for FDA Class II and Medical Device Data Systems approval will be covered.

2018-2019 Precision Care Medicine Projects

Project Name: A Pulse Arrival Time Based Method to Establish Blood Pressure Limits of Autoregulation and Optimal Blood Pressure in Individual Patients During Surgery

Team Members: Yuchen Ge, Shiyu Luo, Bonolo Mathekga, Yinuo Zeng, Shichen Zhang

Advisors: Viachaslau Barodka, MD; Dan Berkowitz; Charlie Brown, MD; Joseph Greenstein, PhD; Raimond Winslow, PhD

Abstract: The regulation of blood flow into peripheral organ systems is imperative to maintain adequate perfusion and prevent tissue damage. The human body accomplishes this task through autoregulation, a mechanism which maintains relatively constant cerebral blood flow despite changes in blood pressure. Autoregulation, however, works only within a limited range of blood pressure and fails to ensure constant cerebral blood flow outside of this range. Under current standards, cerebral oximetry has been used as a surrogate for cerebral blood flow in non-invasive multimodal monitoring of the cerebral autoregulation. However, this method is expensive and thus not widely used in the operating room and intensive care unit (ICU).

In this project, we propose a novel method using real-time continuous electrocardiography and arterial blood pressure data to establish the limits of autoregulation and establish the optimal blood pressure for individual patients in the operating room and ICU. Our method uses a combination of pulse arrival time (PAT), the time it takes the pulse to reach the end organ, and the mean arterial pressure (MAP) to derive the individualized autoregulation limits. Based on the observation that the variance of PAT differs within and outside of the autoregulation limits, the Brown–Forsythe test for constancy of variance is applied between sliding windows across MAP and an adaptive control window, in order to draw the boundaries for the lower and upper limits of autoregulation. Using this method, we are able to determine the autoregulation limits of 100 patients. Future work should include evaluating the performance of our technique against the results of the cerebral oximetry based method.
Project Name: Analytics of Prediction for Insomnia and Cognitive Impairment

Team Members: Ali Al Abdullatif, Amy He, Elysia Chou, John Lin, Jung Min Lee

Advisors: Sridevi Sarma, PhD; Charlene Gamaldo, MD; Rachel Salas, MD; Alyssa Gamaldo, PhD

Abstract: Insomnia is the most common sleep disorder in the United States and affects approximately 60 million Americans. Patients with insomnia also have an increased risk for cognitive impairment. However, insomnia and cognitive impairment are currently diagnosed through subjective clinical interviews with sleep experts and self-reported screening tools. The goal of this study is to create an objective computational model that can predict insomnia and cognitive impairment from physiological recordings obtained during sleep. We construct and test our model using data collected from either polysomnography that was conducted in a sleep center or an ambulatory Sleep Profiler™ device that can be worn at home. Our cohort consists of HIV seropositive patients (n = 31) who are known to suffer from both insomnia and cognitive impairment. Physiological signals from these patients can be used to derive hundreds of sleep stage-specific features, such as average power of a given frequency band from a specific brain area. Generalized Linear Models (GLMs) are then utilized to select the most predictive features for insomnia and cognitive impairment. If successful, our model can provide an accessible and objective means of diagnosing insomnia and identifying patients with cognitive impairment for cognitive behavioral therapy.
Project Name: Physiologic responses to red blood cell transfusion in critically-ill pediatric patients at a university PICU between 2014-2015

Team Members: Mich Fredericks, Andrew Jin, Gaurav Sharma, Jasen Zhang, Roger S. Zou

Advisors: Sridevi Sarma, PhD; Melania Bembea, MD, MPH, PhD;

Abstract: In critical care units across the world blood transfusion decisions currently rely heavily on a patient’s hemoglobin concentration, though there is a consensus that clinical judgment also plays an important role. There is a lack of quantitative data to drive these clinical judgements, causing them to be based more on clinician’s experience than insights from the data. Furthermore, for patients in the Pediatric Intensive Care Unit (PICU), characterization of correlations between the decision to transfuse and effects on physiologic variables and clinically significant outcomes require further investigation.

Our objective is to identify key physiological features that significantly change after transfusion in PICU patients and to ultimately model these changes, and to predict adverse outcomes in order to provide clinicians with more data-driven insights to aid transfusion decisions. In order to accomplish this, we present a 15-month retrospective electronic health record cohort study. We obtained time series data from 2156 pediatric patients admitted to the Johns Hopkins Hospital PICU from July 2014 to October 2015. To remove confounding effects from previous procedures, only the first transfusions in their first PICU visit were analyzed. Preliminary variables investigated include median heart rate (HR), respiratory rate (RR), and peripheral capillary oxygen saturation (SpO2). Transfusions were categorized based on the patient’s pre-transfusion hemoglobin levels in the clinically relevant categories: <5 g/dl, 5-7 g/dl, and ≥7 g/dl. Paired Wilcoxon Ranked Sum test on the 2-hour window comparing before and after transfusion was performed.

Our preliminary analysis determined that median HR significantly decreased after transfusion within 2-hour window, whereas median RR and SpO2 exhibited no significant difference. This study validates our unbiased, exploratory method for statistically identifying physiologic variables that change after transfusion. Initial generalized linear models that predict the post-transfusion states of these three variables have been created taking in a combination of the three variables with demographic information such as age, race, and sex as input variables. The correlation coefficients between the actual and predicted values of these models are 0.8009, 0.8157, 0.6497 respectively. It can be noted that the HR and RR models perform significantly better than the model for SpO2. Additional input variables will be incorporated and further models that predict a greater number of variables will be developed to better predict the physiologic state of the patients as well as the probability of adverse outcomes after transfusion.
Project Name: Quantitative Assessment of TMJD Induced Sleep Disorders and Prediction of Therapy Effectiveness

Team Members: Samana AlGharbi, Archana Balan, Jiaqi Huang, Patrick Myers, Nausheen Tickoo

Advisors: Michael Smith, PhD; Sridevi Sarma, PhD; Joseph Greenstein, PhD; Abhishek Dave

Abstract: Temporomandibular Joint Dysfunction (TMJD) is one of the most common orofacial pain conditions, affecting an estimated 6-12 percent of the US population, primarily women. In addition to pain, the disorder is characterized by sleep disturbance, a major contributor to decreased life quality for patients. TMJD diagnosis is currently highly subjective and is based on three scores which are extracted from patient diaries and questionnaires administered by sleep specialists: Cognitive, Somatic and Pain Catastrophizing scores. This research study split patients into three therapy groups: Behavioral, Cognitive and TMJD education. However, current literature has not yet explored the correlation between the sleep scores and the available therapies. Furthermore, the shortage of certified sleep experts, the subjective nature of the questionnaires, and the absence of objective diagnosis contribute to the improper diagnosis and treatment prescription for TMJD patients. This study provides a quantitative approach to asses TMJD by exploring associated sleep factors as demonstrated by sleep surveys and physiological features extracted from patient polysomnography. Our current investigation has demonstrated correlations as high as 0.75, between specific biological markers and patient sleep scores. These physiological markers can be employed to provide a more quantitative diagnosis for patient TMJD status. They will then be utilized in a predictive model to identify the best therapy type for each patient. The implementation of this model could have significant clinical impact by requiring less specialized physicians to properly prescribe therapy to TMJD patients, eliminating the gap between needed and current sleep specialists.
Project Name: Using Machine Learning Models to Predict the Likelihood of Patient Readmission to ICU

Team Members: Jack Wright, Ryan Hanks, Yang Zhao, Arman Koul, Jiaxin Lin

Advisors: Nauder Faraday, MD, MPH; Adam Sapirstein, MD; Saachin Hebbar; Joseph Greenstein, PhD; Raimond Winslow, PhD; Ran Liu

Abstract: Intensive Care Units (ICUs) cater to individuals with severe injuries and illnesses. Therefore, the patients within suffer from acute anatomic and physiologic derangements requiring constant monitoring and more intense support from hospital staff. Their conditions necessitate rapid diagnosis and intervention of abnormalities to facilitate a period of recovery. A challenge arises, however, in the recognition of sufficient resolution of the pathophysiologic state such that the patient can be safely discharged to a lower intensity environment. The decision to discharge is currently based on the expertise of ICU clinicians, but there is currently no formal method to assist clinicians predict a patient’s chance of success of readmission, so this process is imperfect. As such, ICU readmission rates range from 2-20%. Readmission rates depend on a variety of factors, including demographic characteristics, comorbidities, severity of illness score, duration of index ICU stay, type of ICU, discharge destination, etc. Regardless of what factors led to readmission, a major problem manifests in the rates of in-hospital death for those who are readmitted to the ICU. Compared to patients who are successfully discharged, those who are discharged but return to the ICU are 2-10 times more likely to die in hospital. Unfortunately, current predictive models are insufficiently accurate. In general, these models are built on static parameters and don’t take advantage of the large amount of complex data available in EHR, nor do they take time varying covariates into account. So, generating an algorithm that is able to use all of the available data in order to accurately predict readmission to the ICU would have several important impacts. By improving physicians’ ability to determine resolution of the pathophysiologic state and reduce premature discharge, we could expect morbidity and mortality rates to decrease as well as reduced healthcare costs for patients. Similarly, the information a predictive model provides would allow for better allocation of resources, by letting hospital staff know which ICU’s have higher readmission rates and require more attention. Finally, the features examined during this project could potentially be generalized for use in future predictive models.
Project Name: A Machine Learning-Based Prediction of Cardiac Arrest Outcome Using a Large Multi-Center Database

Team Members: Hanbiehn Kim, Hieu Nguyen, Qingchu Jin, Sharmila Tamby, Tatiana Gelaf Romer, Eric Sung

Advisors: Robert Stevens, MD; Jose Suarez, MD; Christian Storm, MD; Joseph Greenstein, PhD; Raimond Winslow, PhD; Ran Liu

Abstract: Cardiac arrest (CA) poses a significant risk of long-term neurological disability. There remains a large unmet need for accurate and reliable methods to predict post-CA neurological outcomes and mortality. Additionally, no existing CA outcome prediction model utilizes physiological time series (PTS) data, which captures real-time changes in a patient’s health. Several machine learning algorithms (generalized linear model, random forest, XGBoost, and neural networks) were used to compare the predictive performance of models with different feature sets: electronic health records (EHR) alone, PTS alone, and PTS and the combination of EHR and PTS. Ensemble models were established to predict neurological outcome and mortality. Our neurological outcome prediction model achieved a higher sensitivity (0.78), specificity (0.88), and AUC (0.87±0.01) compared to the APACHE baseline standard model (AUC: 0.74±0.01, sensitivity: 0.77, specificity: 0.63). Additionally, our highest performing mortality prediction model (AUC: 0.81±0.01, sensitivity: 0.78, specificity: 0.71) outperforms the APACHE clinical baseline (AUC: 0.75±0.01, sensitivity: 0.86, specificity: 0.56) by 7%. The incorporation of PTS was observed to increase mean AUC by 4% for neurological outcome and 2% for mortality. Our results demonstrate that for both neurological outcome and mortality, our ensemble model performs significantly better compared to the baseline APACHE model. Our results also highlight the value of using PTS in CA outcome prediction models and have allowed us to identify previously overlooked predictive features which merit further investigation. We anticipate that these findings may be used as point-of-care reference for ICU physicians to aid in clinical decision-making.