Registration is Now Closed
Please contact course faculty if you have any questions.
2018.HST.953: Collaborative Data Science in Medicine
HST.953: Collaborative Data Science in Medicine is a course that focuses on the secondary analysis of clinical data that is routinely collected in the process of care. In this fall course, students will work with Boston-area clinicians on research projects with the goal of a publication-ready manuscript at the end of the semester. Several of the papers from last 2 courses have already been published in high-impact clinical journals.
While clinical trials are best in inferring causality, they are not adept at demonstrating small effect size across a population, which is typical given heterogeneity of treatment effect. Moreover, clinical trials typically exclude important subgroups (older patients, those with chronic diseases): findings may not be generalizable to the real-world. Because of the limitations of clinical trials including cost, many practice guidelines are supported by low-quality evidence. To make matters worse, these guidelines are often adopted in countries where funding for research is limited. The digitalization of healthcare data may provides an opportunity to develop locally relevant practice guidelines rather than adopting those that are based on research on populations that may not generalize to. Digital data is proliferating in diverse forms within the healthcare field, not only because of the adoption of electronic health records, but also because of the growing use of wireless technologies for ambulatory monitoring. Since clinical trials may be too expensive to perform in most countries, digital health data provides an opportunity to conduct locally relevant research. Rigorous observational studies have been shown to correlate well with clinical trials across the medical literature in terms of estimates of risk and effect size. The world is abuzz with applications of machine learning in almost every field – commerce, transportation, banking, and more recently, healthcare. These breakthroughs are due to rediscovered algorithms, powerful computers to run them, and most importantly, the availability of bigger and better data to train the algorithms.
A variety of datasets will be available, including MIMIC-III and Philips eICU from the USA. Please complete the Expression of Interest Form to be added to our mailing list and receive email notifications.
There are no prerequisites for this course for MIT, Harvard and Wellesley students. For the rest, we require some experience with R, Python and/or SQL. Everyone is required to complete an online human subjects training (if they haven’t already done so), and sign a Data User Agreement to obtain access to the MIMIC and eICU Collaborative Research Database. This is a project-based course and all the students are required to participate in clinical research using one or both of the databases.
Please, download our syllabus here Syllabus
For more info please, contact us: HST953 Faculty
Fluid resuscitation of patients with CHF
Kyle Whelan/Euma Ishii
One of the cornerstones of sepsis treatment is early fluid resuscitation, but the decision to give fluids is complicated in patients who are more likely to suffer the negative impact of excess fluid administration, particularly in those with congestive heart failure. The aim of this project is twofold: 1) to characterize physician practices with respect to fluid resuscitation in patients with sepsis and a history of known heart failure , and 2) to evaluate outcomes patterns among patients with heart failure and sepsis according to the amount of fluid received. We are using the MIMIC database to identify patients admitted to the ICU at BIDMC with sepsis according to Angus criteria, and within this cohort to compare the amount of intravenous fluid received between patients with and without a prior diagnosis of heart failure. In this way we will elucidate whether physician practice patterns around fluid resuscitation are influenced by a knowledge of prior heart failure. Secondly, among patients identified to have sepsis and some degree of heart failure, we will evaluate the relationship between amount of fluid received in the ICU and incidence of outcomes including mortality, length of ICU stay, and new respiratory failure.
Vancomycin dosing and nephrotoxicity
Vancomycin is an intravenous antibiotic used widely in the treatment of critical illness. The drug’s pharmacokinetics vary based on patient characteristics and disease state, and thus require that vancomycin be monitored by serum testing in order to maintain therapeutic levels. The Infectious Disease Society of America (IDSA) guidelines mandate that these level be tightly controlled between 15-20mg/L in cases of complicated infections in order to ensure therapeutic effect, reduce risk of antibiotic resistance, and prevent iatrogenic harm. Current practice is to base individual patient dosage decisions on standardized nomograms accounting for renal function and body weight. These regimens, however, suffer from low success rates, with published rates of first serum vancomycin level within therapeutic range in only 30-60% of patients. Success rates have been found to be even lower in the critically ill and obese. With the wealth of data becoming available in large clinical datasets, there is now an opportunity to investigate in a much more granular fashion the drivers of vancomycin dose response. Machine learning techniques have shown significant success in assisting clinical decisions in healthcare. More specifically, there have been several examples demonstrating the power of reinforcement learning in deriving optimal dosage strategies from a training set of suboptimal examples. The aim of this proposal is to determine an optimal patient-level vancomycin dosage regimen in order to rapidly achieve and maintain vancomycin levels within a therapeutic range.
Tracking GI bleed using time series data
Daniel Stein/Yuksel Altinel
Gastrointestinal hemorrhage is a common condition that frequently requires hospitalization, often in the ICU. Determination of ongoing bleeding is difficult, since all current methods used are delayed in recognizing the bleeding. We are using ICU vital signs, presenting details and lab work to predict the presence of ongoing bleeding in patients admitted with gastrointestinal hemorrhage. The goal is to develop an algorithm that can predict likelihood of bleeding in real-time to determine which patients need transfusions or interventions such as endoscopy or colonoscopy. We will be determining and validating our model with the MIMIC dataset, comparing the predictive variables to outcomes including interventions, drop in blood concentration (hemoglobin) and transfusion requirement.
Predicting baseline laboratory values
A significant fraction of patients who get admitted do not have medical records available to the providers at the time of admission. When a laboratory test is abnormal, it is important for the providers to determine whether the abnormality is longstanding prior to admission (chronic) or whether it is part of the acute illness. Anemia (low levels of circulating red blood cells) is a common problem and can occur in acute, subacute and chronic forms. The intensity of treatment required (often blood transfusion) depends largely on the chronicicity of the process. However, pre-admission or baseline hemoglobin levels are often not available for critically ill patients admitted to intensive care units. This project will focus on using hemoglobin trends, vital signs, and co-morbidities to back-predict the probable baseline hemoglobin concentration. Kidney injury is a common complication of critical illness. Just like anemia, it is important to determine whether a blood test indicating kidney injury is acute or chronic. The presence of acute kidney injury signifies a greater severity of illness which impacts decisions around treatment. There are patients in MIMIC who receive their outpatient care at BIDMC. For these patients, their baseline hemoglobin and kidney function are available. This creates an opportunity to create an algorithm that predicts the baseline value.
Optimizing treatment of sodium disorders
Hypernatremia is a common clinical entity in patients across all hospital settings. Current best practice of hypernatremia management, in addition to determining the underlying etiology, involves calculating free water deficit and repleting patients with hypotonic fluids while measuring serial serum sodiums. In our anecdotal experience, this method does not always yield expected results, especially among more critically ill patients in the ICU. This experience is supported by studies showing that serum sodium is not corrected at an appropriate rate in the majority of hospitalized patients. Inadequate correction rates are associated with increased mortality,so it appears prudent to determine a more accurate means of predicting sodium changes. In this study, we will look at patients admitted to the ICU with hypernatremia and determine the total free water given over the first 24 hours of admission. This will allow us to ascertain how often current methods of free water repletion lead to appropriate changes in serum sodium in the ICU setting. Our goal will be to understand how various clinical parameters at admission and over the first 24 hours affect the rate of sodium change and ultimately we hope to devise a more reliable prediction model for sodium correction in the first 24 hours of ICU admission.
Hyperoxia in trauma patients
Major trauma is a global problem and comprises approximately 10% of intensive care unit admissions. Oxygen (O2) might the most commonly administered drug in trauma settings, but liberal O2 use and resultant tissue hyperoxia may actually be detrimental to several organ systems. Acute kidney injury (AKI) is a potentially severe complication and very few studies have described the impact of hyperoxemia on the development of AKI in the critically injured. The aim of this project is to investigate whether a liberal use of oxygen can provoke the development of AKI in a trauma setting, and to find a safe therapeutic window for O2 therapy in regards to AKI for patients suffering from severe trauma.
Treatment of status epilepticus in the ICU
Objective: To evaluate whether treatment with continuous infusion (CI) for seizures and status epilepticus (SE) in the intensive care unit (ICU) is associated with higher mortality. Methods: Retrospective descriptive study with data from the Medical Information Mart for Intensive Care III (MIMIC III), a publicly available database of patients in the ICU or with the Phillips Database. We will compare patients who received CI with those who did not. The main outcome will be ICU mortality. The main predictor will be the use of CI of midazolam, propofol, and/or pentobarbital in the ICU. To reduce bias, we will only consider those medications if they will be administered as a continuous infusions and the infusion rate was at least the minimum recommended for SE: 0.05 mg/Kg/hour for midazolam, 30 mcg/Kg/hour for propofol, and 0.5 mg/Kg/hour for pentobarbital. With this approach we aim to minimize the proportion of CI given for sedation, intubation, or other purposes. Variables that may potentially influence the probability of receiving CI and/or the probability of death will be: age, gender, ethnicity, insurance, admission origin (from the emergency room, other hospital, etc.), sequential organ failure assessment (SOFA) score, central nervous system function (measured with the Glasgow coma scale), ICU setting (medical ICU, coronary unit, etc.), and presence of the following chronic conditions: congestive heart failure, cardiac arrhythmias, valvular disease, pulmonary circulation disease, peripheral vascular disease, hypertension, paralysis, neurological disease, chronic pulmonary disease, uncomplicated diabetes, complicated diabetes, hypothyroidism, renal failure, liver disease, AIDS, lymphoma, metastatic cancer, solid tumor, rheumatoid arthritis, coagulopathy, obesity, weight loss, fluid-electrolyte imbalance, anemia or blood loss, deficiency anemia, alcohol abuse, drug abuse, psychosis, and depression.
Trial of ICU for elderly patients with septic shock
The prognosis for critically ill patients with a severe infection complicated by sepsis is poor. Time limited trials of aggressive care in intensive care units (ICU) are often used in these patients to assess for possible improvement with aggressive treatment. However, the optimal length of such trials is unknown. The goal of this project is to build a decision model based on Markov microsimulations to identify the optimal duration of intensive care for critically ill patients with sepsis. The MIMIC-II database will be used to calculate state transition probabilities between different states (clinical improvement, clinical deterioration, transfer out of the ICU, discharge from the hospital, death) conditional on the patient’s severity of illness using Kaplan-Meier time-to-event analysis. Different lengths of time limited aggressive ICU care will be compared based on the 30-day and mean overall mortality. The primary outcome of interest is the length of aggressive care in days after which additional intensive treatment does not significantly improve survival anymore.
We will use MIMIC and eICU to quantify the effect of IV fluid administration (principally normal saline, though will explore other crystalloids as time permits) on hemoglobin and hematocrit, routine blood tests which describe the concentration of red blood cells in a given patient. Reliably trending hemoglobin and hematocrit is of paramount importance in numerous clinical scenarios (such as in patients who are bleeding), and it is generally assumed that IV fluid administration has a diluting effect on these lab tests by increasing intravascular plasma volume and consequently decreasing red blood cell concentration. This concept is called “hemodilution”. While hemodilution intuitively seems believable enough, there is no definitive evidence that such a trend exists, nor is there any agreement among clinicians as to the size of the effect. The potential harm in belief in this unproven effect is clear: for instance, a patient with internal bleeding may be delayed from receiving life-saving surgery if the physicians believe that his or her drop in blood concentration is due to volume expansion rather than true blood loss.
This project will build off of extensive work already performed over the past 9 months, in which we have used selected cohorts from MIMIC and eICU to demonstrate that there is a strong negative correlation (p < 0.0001) between IV normal saline and hemoglobin and hematocrit levels. Our primary goal is now to specifically quantify the expected drop in hemoglobin and hematocrit given various volumes of IV fluid. Secondary goals include repeating this analysis with different crystalloids or colloids and evaluating the presence and strength of effect on other lab tests such as platelet count and white blood cell count.
Efficiency of Delivering Intensive Care Resources Over Time
Jesse Raffa/Shawn Sturland/Laura Myers/Ben Geisler
Using MIMIC data, we will work with multiple groups of students to describe how clinical outcomes have changed between 2002-2012 for various diagnoses (CAP, sepsis, renal failure, GIB) receiving specific interventions (mechanical ventilation, pressors, renal replacement therapy, blood transfusion therapy, respectively). We believe that use of these interventions has gone down over time yet in-hospital and 1-year mortality have improved. We will need to account for correlation of the data over time, as well as multiple strong confounders such as secular trends, severity of illness, and hospice status, for example. Depending on the interests of the groups, we may be able to examine differences in delivery of the above interventions between hospitals using Philips eICU data, as well as the speed at which seminal practice changes were implemented using MIMIC data. Decision Support to Estimate Benefit of Dialysis in Sepsis with Kidney Injury In a study involving 54 hospitals from across 23 countries including both rich and poor countries, up to 2⁄3 of patients admitted to the intensive care unit (ICU) developed acute kidney injury or AKI. Over-all, 4-5 of ICU patients require dialysis, a proportion that was surprisingly consistent across the countries. 50-70% of patients who required dialysis in the ICU die, again a proportion that is surprisingly consistent across both rich and poor countries. This means that in over half of the patients receiving dialysis, the procedure does not afford survival benefit and represents a wasted very expensive resource. A decision support tool that predicts whether a patient will benefit from dialysis if they develop severe kidney injury will be quite helpful.
Machine Learning to Identify Symptoms from the Notes
In this project, we will use machine learning methods (e.g., CRFs, NeuroNER) to extract documentation of heart failure symptoms from the EHR. The team will have access to already developed annotation software to generate the training dataset. As a second step, we will correlate chest x-ray images and structured data to symptom documentation. The project will explore how well machines can identify patient symptoms without involving a clinician.
Detecting Bias and Prejudice from the Notes: Alcoholic vs. Non-alcoholic Liver Disease, Sepsis with and without Morbid Obesity
There are 3 major causes of end-stage liver disease in the US: alcohol abuse, hepatitis C (which is typically obtained from intravenous drug use) and non-alcoholic steato-hepatitis or NASH (a condition that is neither from alcohol or other known causes of liver disease). The project will compare the notes written by nurses, doctors and social workers to document their care for these patients. Is there a difference in the way the patients are described by their providers based on whether the liver disease is from alcohol abuse vs. NASH?
Using Real-World Data to Validate or Refute Physiologic Dogmas
Sarah Train/Leo Anthony Celi
Numerous physiologic concepts have been demonstrated in animal models without specific validation in humans. Even more limited is the understanding of how these concepts change in the critically ill patient, where patients are exposed to extreme physiologic stresses. Using the MIMIC III database, we are able to explore how the physiology of critical illness changes (or doesn’t) these widely-accepted principles. For instance, we are working on using MIMIC III to explore the impact of critical illness on the shape of the oxygen dissociation curve (ODC). Using arterial blood gas measures, complete ODCs have been generated for these patients. We hope to use this information to identify and characterize patients with ODCs that deviate significantly from the expected curve.
We hope to continue to explore various physiologic concepts in this manner to better understand the changes in physiology that occur with critical illness.
Heparin dose optimization
The current dosing guidelines for unfractionated heparin (UFH) are weight-based, established on the results of a randomized control trial in 1993 demonstrating more rapid achievement of therapeutic anticoagulation compared to a standard care nomogram. Unfortunately, when using a weight-based formula in critically ill patients there are significant delays in achieving therapeutic anticoagulation. Subtherapeutic anticoagulation fails to adequately treat the condition for which it is indicated, while supratherapeutic anticoagulation places patients at risk of hemorrhage. Prior research efforts using the Multi-parameter Intelligent Monitoring in Intensive Care database (MIMIC-II) used logistic regression to create a predictive model for achieving therapeutic anticoagulation with UFH. We propose to expand on the work done using MIMIC-II by applying machine learning techniques to the Medical Information Mart for Intensive Care database (MIMIC-III) to identify and incorporate novel predictors into a dosing model for UFH that will improve the time spent with therapeutic anticoagulation.