David A. Clifton, Kellogg Junior Research Fellow Institute of Biomedical Engineering University of Oxford
The clinical need Method: machine learning for vital-signs monitoring Results and clinical trials so far
The clinical need Method: machine learning for vital-signs monitoring Results and clinical trials so far
23,000 preventable cardiac arrests occur every year in UK hospitals 20,000 unplanned ICU readmissions every year Between 5% and 24% survival rate The majority of these occur because physiological deterioration goes undetected... Why?
Level 3:ICU 1:1 Level 2: Step-down 1:4 Level 1: Acute wards 1:4 Level 0: General wards 1:10 Level -1: Home 1:? Existing patient monitors generate very high numbers of false alerts (e.g. 86% of alerts in 1997 MIT study)
NPSA, Safer Care for Acutely Ill Patients (2007) Failure to identify: 73% of reviewed patient deaths are preceded by deterioration that is not identified or acted upon NICE, Acutely Ill Patients in Hospital (2007) Objective monitoring: Recommended aggregate weighted scoring systems (track-and-trigger) in all UK hospitals: HR, BR, SpO2, Systolic BP, Temp, Consciousness
1615 observations: 72 (4.5%) have all scores 242 (15.2%) have some scores 1297 (80.3%) have no scores (HR, BR, SpO2, SysBP, DiasBP, GCS, Urine, Temp) 314 observations with scores: 310 with scores totalled correctly 2 with scores totalled incorrectly 2 with scores not totalled (19.2% of total) (0.1% of total) (0.1% of total)
Summary of existing state-of-the-art: Patients need monitoring, because the nurse:patient ratio decreases from 1:1 outside ICU Existing bed-side patient monitors have 86% false-positive rate, and are ignored Recommended track-and-trigger systems are incomplete, with no evidence base (heuristic) Therefore: intelligent patient monitoring, with reliable false-alarm rates
The clinical need Method: machine learning for vital-signs monitoring Results and clinical trials so far
Heart rate (ECG + Oximetry) Breathing rate (ECG + IP) SpO2 (Oximetry) Blood pressure (Cuff) Temperature (Cuff)
EEG Heart rate (ECG + Oximetry) Breathing rate (ECG + IP) SpO2 (Oximetry) Blood pressure (Cuff) Temperature (Cuff)
PSI Patient Status Index (alarm > 3.0) Vital-sign data
Heart rate Breathing rate SpO2 Blood pressure Temperature Training Population Density estimation
This model of normality p(x) is stored in Visensia s memory When Visensia is used to monitor a high-risk patient, an alert is generated whenever the vital signs are about to go outside the boundaries of normality Outside the boundaries of the normal data, p(x) will have a very low value. 1. Define patient status index : PSI = log e [1 / p(x)] 2. Set threshold for alerting at PSI = 3.0 (on theoretical grounds)
Extreme Value Theory (EVT) characterises the tails of distributions Determines our expectation of where extrema generated from some pdf, p(x), will lie allows us to determine where to set the alarm threshold Previously proven for health monitoring of jet engines (using engine pressures, engine temperatures, etc.)
The clinical need Method: machine learning for vital-signs monitoring Results and clinical trials so far
1. John Radcliffe Hospital (Oxford) 440 high-risk elective/emergency surgery or medical patients September 2003 to July 2005 2. Clarian Methodist (Indianapolis) 220 patients from upper end of general floor or Progressive Care Unit (PCU) January 2006 to June 2007 3. University of Pittsburgh Medical Center (UPMC) 1,000 patients from 24-bed Step-Down Unit November 2006 to August 2007
Hravnak et al, Archives of Internal Medicine (2008) Medical Emergency Team (MET) criteria fulfilled due to cardiorespiratory events occurring on 7 occasions All 7 events were detected by PSI in advance (mean advanced detection time prior to fulfilment of MET criteria was 6.3 hours) Cardiopulmonary deterioration was usually characterised by progressive increases in the Visensia Index over time, not step increases
There were 0.94 false alerts per 100 hours of monitoring This corresponds to a false alert rate of 0.23 per patient per day. The Visensia data fusion model automatically switches to a lower-dimensional model when a parameter is artifactual or missing This makes the technology usable by the nursing team
Hravnak et al, MET Conference, Toronto (2009) Three-fold reduction in the number of patients becoming critically unstable for a sustained period of time (17.8% in Phase 1, 5.2% in Phase 3) Data fusion system was not withdrawn from the SDU at the end of the 6-month trial No unexpected fatal cardiac arrests in last 18 months (compared with 50 in previous 18 months, prior to introduction of data fusion technology)
Emergency Dept., Oxford JR Haemodialysis, Oxford Churchill Cancer Hospital, Oxford Churchill HotF, Oxford JR Guy s & St. Thomas, London St. Joseph Mercy Oakland
Wireless sensors... acquire data from mobile patients, and transmit the results via Bluetooth and Wifi to Visensia devices
Patients monitored in bed, then monitored remotely, as they are allowed to wander Patients in distributed hotel rooms now trackable via Nursing Central Station
Intelligent patient vital-sign monitoring has seen excellent preliminary results Rapid realisation of benefit for patients and hospital staff Extension throughout the hospital, to all levels of care Ambulance-to-home monitoring, with a Visensia addition to the electronic patient record? Department Status Index for managing clinical resources Plenty of scope for further statistical and machine learning contributions
David A. Clifton, Kellogg Junior Research Fellow Institute of Biomedical Engineering University of Oxford