Suicide is the tenth leading cause of death and alarmingly, suicide rates have increased over the past two decades. One reason for the limited progress in suicide prevention is that clinicians lack a strong ability to predict when and for whom suicidal behavior will occur.
Our goal is to leverage advances in digital monitoring technology and personalized machine learning to improve suicide prediction. One behavioral marker of suicide risk that has received increasing attention is speech production. As most individuals who die by suicide deny suicidal intent during their last communication with providers, objective speech features reflecting how one speaks (e.g., pitch inflection, phoneme specific variance, prosody) may have greater predictive utility than speech content. One limitation of prior efforts, however, is that speech has been assessed in highly structured formats and at a single time point, which limits ecological validity and prediction accuracy of the temporal window of future events.
We propose to longitudinally monitor the unstructured speech of adults hospitalized for suicidal ideation or attempt on the MGH psychiatric inpatient unit. Participants in an ongoing digital monitoring study (the “parent study”) will additionally record daily audio diaries via smartphone. We will train personalized machine learning algorithms using voice characteristics and other clinical and real-time data to prospectively predict withinindividual changes in suicidal ideation and behavior on the unit and post-discharge (the highest risk period for suicide). We will also assess incremental predictive power afforded by speech features, beyond other data. Results will inform clinically applicable suicide prevention efforts.
Ischemic stroke is the fifth leading cause of death in the United States. Of the 800,000 Americans yearly who experience a stroke, approximately 185,000 are recurrent attacks. While significant work has been conducted to determine relevant risk factors, uncertainty remains regarding 1) a single individual's relative risk of suffering an additional stroke, and 2) how certain subgroups will respond to medical management strategies like antiplatelet and anticoagulation. In order to improve personalized risk assessment, we need to harness our ability to employ predictive pattern technology and integrate computer assisted analysis of brain/vascular imaging with clinical data. The author’s preliminary work on radiographic markers of stroke patients demonstrate that discrete imaging components that significantly predict future outcomes. Our main objective for the MIT Philips Research Award is to use novel methods to automate the analysis of radiographic data and use it to construct personalized risk predictions and medical management recommendations for patients with ischemic stroke. We will then couple it to clinical data to construct an enhanced prediction tool that determines an individual's risk of further stroke and identifies optimal medical therapy. The proposed research is innovative because we propose to develop versatile, and comprehensive imaging analysis tools using novel machine learning approaches applicable to the study of personalized secondary stroke prevention. It will have significant impact on the population because it can provide clinicians and researchers with a new array of strategies to optimize treatment plans for secondary stroke prevention, decreasing the burden of stroke in the population.
Over 500,000 cardiac arrests happen yearly in the United States. No available monitoring tool provides quantifiable and real-time feedback about the function of the neural networks associated with emergence from coma after hypoxic brain injury. We propose to evaluate the performance of machine learning methods that utilize quantitative electroencephalogram (qEEG) for hypoxic coma outcome prediction. Our preliminary work indicates that dynamic changes on EEG signals to environment stimuli, i.e. EEG background reactivity (EBR), strongly predicts poor outcome. To test if a computational method can detect EBR, we have developed a qEEG toolbox (qEBR) that outperformed human raw EEG analysis using non- continuous EEG data. We hypothesize that machine-learning techniques employing qEBR in large EEG datasets will: a) have high accuracy predicting functional recovery, b) be scalable to qEEG data obtained exclusively from forehead electrodes, and c) adapt to the effects of temperature and anesthetics on EEG signal properties. This data-driven decision support system has the potential to identify early signs of neurological improvement that precede awakening from coma, and therefore encourage intensive supportive care and promote survival after cardiac arrest. Additionally, a high performance portable neuromonitoring application may increase access and reduce costs of brain physiology monitoring beyond the intensive care unit environment. The innovation of this project lies in 1) application of novel machine learning methods to identify brain recovery biomarkers, 2) development of multi-channel signal processing techniques applicable to wearable devices, 3) design of context-sensitive prediction models that integrate multiple streams of physiology data.
Tight oxygen titration in the preterm neonate is a key aspect of neonatal intensive care due to the mortality associated with hypoxia and the morbidity associated with hyperoxia in this vulnerable population. Despite these known complications of sustained oxygenation outside the corrected gestational-age target ranges, most Neonatal Intensive Care Units (NICUs) fail to reliably maintain infants’ saturations within target range.
The primary goal of this project is to identify clinical, demographic, physiological and workflow factors that place preterm infants at risk for hypoxia and hyperoxia. To achieve this goal, we will leverage the copious Philips bedside monitoring data archived from our NICU under the Multiparameter Intelligent Monitoring in Intensive Care II (MIMIC-II) project, and going forward using the Philips Data Warehouse Connect system. By identifying oxygenation patterns from the bedside monitoring data and relating such patterns to clinical, demographic, physiological and workflow factors, we are confident that our work will lead us to identify specific and generalizable strategies to optimize oxygenation in the preterm population. Modifiable factors will directly feed into our ongoing Quality Improvement campaign for rapid- cycle assessment of improvements in in-target-range oxygenation.
More broadly, by bringing together the combined expertise in Quality Improvement science (BIDMC Department of Neonatology), critical care data analytics (Professor Heldt’s laboratory at MIT), and patient monitoring (Philips Healthcare), this project is uniquely positioned to demonstrate how closing the loop on ‘big-data’ analysis can directly inform and rapidly transform clinical care for demonstrable and quantifiable improvement in outcome.
Renal recovery after acute kidney injury (AKI) is associated with a higher risk of death and adverse renal outcomes after hospital discharge. To better understand the long-term consequences and accurately describe the epidemiologic burden of AKI, more refined definitions of AKI are needed.
Specific Aim 1: To determine the impact of AKI on long term mortality when excluding ICU mortality.
Hypothesis: AKI on admission will be associated with increased mortality, with a stronger association closer to the time of discharge and a weaker association farther from discharge.
Specific Aim 2: To determine the impact of AKI recovery on long term survival.
Hypothesis: Lack of renal recovery at discharge will have a stronger association with mortality compared with those with recovery.
The outcomes of the proposed studies could identify important methodologies in risk stratification of patients with AKI. This can then be used to build additional methods of risk categorization in patients with AKI in the ICU.