Methods
Funding for this study was provided through ‘Fighting Ebola: A Grand Challenge for Development’, launched by the US Agency for International Development (USAID) in partnership with the White House Office of Science and Technology, the Centers for Disease Control and Prevention, and the Department of Defense. It was carried out in collaboration with the humanitarian organisation International Medical Corps.
ETC—Makeni, Sierra Leone
An ETC opened outside Makeni, Sierra Leone in December 2014, funded by the UK's Department for International Development and run by International Medical Corps was the site of this study6 (see online supplementary figure S1). It was a 50-bed facility that provided care 24/7 by national and international workers distributed into three primary teams based on principal responsibilities: Medical, Water, Hygiene and Sanitation (WASH) and Psychosocial.
Individuals suspected of potentially having Ebola based on their contact history, presence of a fever, signs of easy bleeding and systemic symptoms were transferred to the ETC and evaluated in a triage zone for admission. Individuals meeting admission criteria, who were aged 18 years or older, were able to understand the informed consent, and without a prior history of skin sensitivity to adhesives, were approached for enrolment and informed consent was obtained. The informed consent document was written in English, the official language of Sierra Leone; however, for the majority of individuals, a member of the medical team explained each component of the informed consent to them in their local language to ensure full understanding.
On the day of admission and 24 hours later, all patients were tested for Ebola virus in their bloodstream by quantitative PCR. After two negative tests, and if clinically stable, patients were discharged. For anyone testing positive, they were transferred to the confirmed ward of the ETC and remained hospitalised until they recovered enough to be discharged, which was on average ∼16 days. All individuals also underwent rapid diagnostic testing for malaria, which was reported as either positive or negative, although all admitted patients were empirically treated with a combination antimalarial agent (Coartem) during their admission to the ETC. In addition, all admitted patients received empiric antibiotics, antipyretics and oral or intravenous rehydration based on their volume output and degree of dehydration.7
Medical team members in full PPE manually obtained vital signs at the time of admission and then routinely three times a day. Blood pressure values were never measured. Temperature was obtained by infrared thermometer, pulse by radial artery palpation and respiratory rate by observation.
Personalised physiology analytics
For data extraction, storage, analysis, visualisation and initial development of a personalised, predictive, automated alerting tool, we used the physIQ (http://www.physiq.com) data analytics platform, known as PPA for personalised physiology analytics. Their technology is based on machine learning algorithms that are purely data driven and can be applied to a wide range of vital signs parameters. Using the biosignal waveform data from any wearable sensor set, physIQ feature extraction algorithms generated the following parameters at 1 min intervals: heart rate, heart rate variability (both time domain and spectral analysis), activity, respiratory rate, pulse transit time (PTT; inversely related to blood pressure), posture, ECG quality metric and arrhythmia burden. Data quality is based on the ECG signal since many of the vital signs are derived from the ECG waveform. An algorithm is applied to the ECG signal that assesses the ECG information signals strength, motion artefact content and noise levels over 1 min windows of the data. This results in a quality metric that ranges from 0 to 100% for each min data window. If the metric falls below 75%, the corresponding 1 min data window is rejected for use in calculating vital signs from the raw biosignal data.
PPA, cleared by the US Food and Drug Administration in June 2015, is used to detect subtle changes in an individual's physiological characteristics from learnt baseline physiological behaviour. This is accomplished by detecting changes in the inter-relationships between parameters, rather than just examining parameters individually relative to a population norm or statistic. Using machine learning technology known as similarity-based modelling (SBM) each patient's current vital signs are compared with his or her unique baseline in real time to develop a series of residuals.8 ,9 These residuals are then fused into a single index—the Multivariate Change Index (MCI). The MCI represents the likelihood of change in the patient's physiology over time relative to the learnt baseline, identifying the earliest signs, often presymptomatic, of an improving or worsening medical condition that can be graphically presented as a time trend to healthcare workers. The baseline physiology learnt by SBM for each patient was based on vital signs data collected during the first day of monitoring to capture a full diurnal cycle of physiological activity. As it was unknown at the time of admission the clinical trajectory the patient would follow a version of MCI was applied that generates a signed likelihood of change where the magnitude is the likelihood of change and the sign is indicative of improved health conditions (positive) or degraded health conditions (negative). The sign is driven by the patterns produced across the multiparameter SBM residuals at each instant in time. If the monitored vital signs are behaving similarly to the baseline, the residual values will all be close to zero and the MCI will be close to zero, indicating no change in physiology. However, if the monitored vital signs begin to deviate from the baseline, one or more of the residuals will increase in magnitude and will be biased either positive or negative relative to the amount of change seen in each vital sign from the baseline period. This pattern of positive, negative and near-zero residual components forms the patterns that drive the sign of MCI. A detailed explanation of MCI and determination of baseline deviation can be found in ref. 9
Wearable sensor
For this proof of concept project, we used the MultiSense patch being developed by Rhythm Diagnostics Systems (http://www.rhythmdiagnosticsystems.com). The MultiSense is a self-contained, battery-powered, flexible strip, measuring 4×1.2 inches and weighing <15 g (see supplementary figure S2). It is waterproof with a patented adhesive for reliable adherence to the chest in most situations. It is able to continuously and simultaneously track and store the following biosignal waveform data: ECG (sampled at 256 Hz), three-dimensional accelerometer (16 Hz), uncalibrated skin temperature (4 Hz), red and infrared photoplethysmography (PPG; 32 Hz). For this study, the majority of the patches used were first-generation, memory-only devices that stored all physiological data in on-board memory and could be worn for up to 10 days with the data then downloaded via Universal Serial Bus (USB) once removed. To demonstrate its efficacy for the continuous, real-time streaming of vital signs, the second-generation MultiSense was used for two patients. This device is otherwise identical to the first-generation device with the addition of being Bluetooth-enabled. For this project, an Android device running the physIQ app, without cellular connectivity but with Bluetooth and Wi-Fi, was used to collect data from the patch and transmit it from the contaminated area (red zone) to the clean area (green zone) where personal protective gear was not required. The Android device was maintained by the patient's bedside.
Portable, deployable monitoring infrastructure
Basic information technology (IT) infrastructure was limited in the Makeni ETC, as it would be expected to be in many settings requiring a rapid medical response. As a result, one of the fundamental requirements of this project was to engineer and produce a portable, self-contained ‘system in a trunk’ providing all required IT infrastructure except power. The MWPMS needed to be readily shipped and deployable in remote locations without support capabilities. The end result was a system designed to fit into a Transportation Safety Administration-approved container that could be checked as baggage on a commercial airline (see online supplementary figure S3). The prototype system contained a hardened Panasonic Tough Book laptop, preloaded and configured to run the physIQ PPA platform for remote patient monitoring.
For this project, two communication systems were provided. The first was a dedicated private Wi-Fi network that provided communication for the wireless patient-worn devices and for the tablets used by the clinicians. A second satellite data link was provided for remote system administration that eliminated the need for onsite technical staff. For future projects, the satellite link would also allow for the patient population to be securely monitored by clinical staff throughout the world.
The laptop was connected to a hardened self-contained network hub that provided the entire network routing and switching for the system. From the network hub, the Cat6 Ethernet cable was run up to 300 ft to a weatherproof Wi-Fi radio attached to a high gain antenna. The radio and antenna combination provided a Wi-Fi network that covered the entire Ebola Treatment Unit (ETU) when centrally located. Power over Ethernet was used to power the radio from the network hub, eliminating the need for any additional power cabling to be run to the radio. The satellite link used a Hughes 9502 Integrated BGAN radio, which allowed physIQ staff in Naperville, Illinois, USA, to remotely monitor and maintain the system in Sierra Leone.
Since sensor data were received over the Wi-Fi network, they were stored in a time-series database on the Tough Book and processed for clinician review by the physIQ PPA platform. Via the two provided Wi-Fi enabled tablets, clinicians could then view, while in the green zone, the vitals from the patients as well as the trending MCI.
Statistical analysis
Descriptive statistics were used to describe demographic and clinical characteristics such as coinfections. Pearson's correlation coefficients were determined for relationship between continuously sensed and manually determined vital signs. Manually measured vital signs included heart rate (peripheral pulse), respiration rate (by observation) and temperature. The time stamps for manually measured vital signs where assigned after patient rounds were completed, and could be off from the actual time of measurement by 20 min or more. To compare to the continuous vital signs measurements, trimmed mean averages of the continuous vital signs were calculated over a time window of ±20 min from the reported manually measured vital signs time stamp. The trimmed mean was calculated by taking the average of the continuous vital signs data in a 40 min window after removing the upper and lower 10% of the samples. Some manually measured vital signs occurred outside the period of actual continuous monitoring. Those samples were ignored. The remaining manual vital signs samples and continuous vital signs averages were analysed using Pearson's correlation coefficient over the patient population. In the case of temperature, since the continuous skin temperature measurement is uncalibrated, the difference between the mean of the manually measured temperatures and the continuously measured temperatures for each patient was calculated and used to offset the uncalibrated readings before applying the Pearson's correlation coefficient analysis to temperature.