Optimization of monitoring, prediction and phenotyping of deterioration of inhospital patients using machine learning and multimodal real time data

点击翻译按钮获取中文摘要

Efficient patient monitoring on the medical-surgical wards is crucial, because up to 5% of hospitalized adult patients deteriorate, requiring transfer to the intensive care unit (ICU) or intervention of a rapid response team (RRT). Currently, vital sign measurement is performed on all patients every 4-6 hours, even the most stable. For stable patients, this monitoring is often unnecessary, whereas for higher-risk patients, vital sign monitoring every 4-6 hours is often not adequate. To address this need, we will leverage one of the largest, most diverse clinical datasets in the country, using electronic health record (EHR) data from 2.4M hospitalized patients to generate machine learning (ML) predictive models, designed to optimize patient monitoring. We will use continuous monitoring (CM) devices to identify in advance patients likely to deteriorate and specify the clinical underlying reasons of deterioration to enable timely interventions. We have applied and published similar ML approaches on other cohorts, including: 1) deep recurrent neural networks (RNNs) to avoid unnecessary overnight vitals; 2) deep learning models that use continuous monitoring data to predict clinical alerts up to 4 hours ahead of time; and 3) natural language processing on medical notes and unsupervised clustering of patients. Our approach involves collecting prospectively CM data from a targeted population of 2,000 hospitalized patients, and developing and validating models, both retrospectively and prospectively. Our approach will allow us to: Identify stable patients admitted on the medical-surgical wards to optimize vital signs monitoring. We will train a RNN model using EHR data from 2.4M hospitalizations, to predict, after vital signs are measured, stable patients for the next 8 hours, and enable eliminating the next vitals measurement. We retrospectively will validate the model, using cross-affiliation validation, and prospectively, silently validate it in 5 different hospitals. Develop a clinical deterioration algorithm, based on continuous monitoring data and clinical hard outcomes. We will collect prospective data from a targeted population of 2,000 inpatients, who are admitted on medical-surgical floors in our largest hospital, with a modified early warning score higher than 5. The CM patches will start collecting data upon admission. We will use combined clinical hard outcomes (death, intubation, cardiac arrest, unplanned ICU transfer, RRTs) to train two deep- learning models to predict deterioration up to 4 hours and up to 24 hours before. Define the early and late phenotypic substrates of hospitalized patient deterioration. Using the clinical data of 56K deteriorated patients from Aim 1 (EHR variables and extracted presenting symptoms) 4 hour and 24 hours prior to deterioration, we will perform unsupervised cluster analysis to identify unique clusters linked to phenotypes of deterioration. We will associate derived phenotype groups to clinical outcomes and treatments, to inform more targeted treatment and intervention strategies. We aim to develop new tools to align patient needs with resources, and deliver more efficient, effective, personalized, and proactive care to hospitalized patients.