Machine learning to refine decision making within a syndromic surveillance service.

BACKGROUND: Worldwide, syndromic surveillance is increasingly used for improved and timely situational awareness and early identification of public health threats. Syndromic data streams are fed into detection algorithms, which produce statistical alarms highlighting potential activity of public health importance. All alarms must be assessed to confirm whether they are of public health importance. In England, approximately 100 alarms are generated daily and, although their analysis is formalised through a risk assessment process, the process requires notable time, training, and maintenance of an expertise base to determine which alarms are of public health importance. The process is made more complicated by the observation that only 0.1% of statistical alarms are deemed to be of public health importance. Therefore, the aims of this study were to evaluate machine learning as a tool for computer-assisted human decision-making when assessing statistical alarms. METHODS: A record of the risk assessment process was obtained from Public Health England for all 67,505 statistical alarms between August 2013 and October 2015. This record contained information on the characteristics of the alarm (e.g. size, location). We used three Bayesian classifiers- naïve Bayes, tree-augmented naïve Bayes and Multinets - to examine the risk assessment record in England with respect to the final 'Decision' outcome made by an epidemiologist of 'Alert', 'Monitor' or 'No-action'. Two further classifications based upon tree-augmented naïve Bayes and Multinets were implemented to account for the predominance of 'No-action' outcomes. RESULTS: The attributes of each individual risk assessment were linked to the final decision made by an epidemiologist, providing confidence in the current process. The naïve Bayesian classifier performed best, correctly classifying 51.5% of 'Alert' outcomes. If the 'Alert' and 'Monitor' actions are combined then performance increases to 82.6% correctly classified. We demonstrate how a decision support system based upon a naïve Bayes classifier could be operationalised within an operational syndromic surveillance system. CONCLUSIONS: Within syndromic surveillance systems, machine learning techniques have the potential to make risk assessment following statistical alarms more automated, robust, and rigorous. However, our results also highlight the importance of specialist human input to the process.

Developing influenza and respiratory syncytial virus activity thresholds for syndromic surveillance in England.

Influenza and respiratory syncytial virus (RSV) are common causes of respiratory tract infections and place a burden on health services each winter. Systems to describe the timing and intensity of such activity will improve the public health response and deployment of interventions to these pressures. Here we develop early warning and activity intensity thresholds for monitoring influenza and RSV using two novel data sources: general practitioner out-of-hours consultations (GP OOH) and telehealth calls (NHS 111). Moving Epidemic Method (MEM) thresholds were developed for winter 2017-2018. The NHS 111 cold/flu threshold was breached several weeks in advance of other systems. The NHS 111 RSV epidemic threshold was breached in week 41, in advance of RSV laboratory reporting. Combining the use of MEM thresholds with daily monitoring of NHS 111 and GP OOH syndromic surveillance systems provides the potential to alert to threshold breaches in real-time. An advantage of using thresholds across different health systems is the ability to capture a range of healthcare-seeking behaviour, which may reflect differences in disease severity. This study also provides a quantifiable measure of seasonal RSV activity, which contributes to our understanding of RSV activity in advance of the potential introduction of new RSV vaccines.

Monitoring telehealth vomiting calls as a potential public health early warning system for seasonal norovirus activity in Ontario, Canada.

Norovirus is a predominant cause of infectious gastroenteritis in countries worldwide [1-5]. It accounts for approximately 50% of acute gastroenteritis (AGE) and >90% of viral gastroenteritis outbreaks [6, 7]. The incubation period ranges between 10 and 48 h and illness duration is generally 1-3 days with self-limiting symptoms; however, this duration is often longer (e.g. 4-6 days) in vulnerable populations such as hospital patients or young children [2, 8]. Symptomatic infection of norovirus presents as acute vomiting, diarrhoea, abdominal cramps and nausea, with severe vomiting and diarrhoea (non-bloody) being most common [2, 5, 9].

Estimating the burden on general practitioner services in England from increases in respiratory disease associated with seasonal respiratory pathogen activity.

Understanding the burden of respiratory pathogens on health care is key to improving public health emergency response and interventions. In temperate regions, there is a large seasonal rise in influenza and other respiratory pathogens. We have examined the associations between individual pathogens and reported respiratory tract infections to estimate attributable burden. We used multiple linear regression to model the relationship between doctor consultation data and laboratory samples from week 3 2011 until week 37 2015. We fitted separate models for consultation data with in-hours and out-of-hours doctor services, stratified by different age bands. The best fitting all ages models (R2 > 80%) for consultation data resulted in the greatest burden being associated with influenza followed by respiratory syncytial virus (RSV). For models of adult age bands, there were significant associations between consultation data and invasive Streptococcus pneumoniae. There were also smaller numbers of consultations significantly associated with rhinovirus, parainfluenza, and human metapneumovirus. We estimate that a general practice with 10 000 patients would have seen an additional 18 respiratory tract infection consultations per winter week of which six had influenza and four had RSV. Our results are important for the planning of health care services to minimise the impact of winter pressures. •Respiratory pathogen incidence explains over 80% of seasonal variation in respiratory consultation data.•Influenza and RSV are associated with the biggest seasonal rises in respiratory consultation counts.•A third of consultation counts associated with respiratory pathogens were due to influenza.

The burden of seasonal respiratory infections on a national telehealth service in England.

Seasonal respiratory illnesses present a major burden on primary care services. We assessed the burden of respiratory illness on a national telehealth system in England and investigated the potential for providing early warning of respiratory infection. We compared weekly laboratory reports for respiratory pathogens with telehealth calls (NHS 111) between week 40 in 2013 and week 29 in 2015. Multiple linear regression was used to identify which pathogens had a significant association with respiratory calls. Children aged <5 and 5-14 years, and adults over 65 years were modelled separately as were time lags of up to 4 weeks between calls and laboratory specimen dates. Associations with respiratory pathogens explained over 83% of the variation in cold/flu, cough and difficulty breathing calls. Based on the first two seasons available, the greatest burden was associated with respiratory syncytial virus (RSV) and influenza, with associations found in all age bands. The most sensitive signal for influenza was calls for 'cold/flu', whilst for RSV it was calls for cough. The best-fitting models showed calls increasing a week before laboratory specimen dates. Daily surveillance of these calls can provide early warning of seasonal rises in influenza and RSV, contributing to the national respiratory surveillance programme.

Developing and validating a new national remote health advice syndromic surveillance system in England.

Background: Public Health England (PHE) coordinates a suite of real-time national syndromic surveillance systems monitoring general practice, emergency department and remote health advice data. We describe the development and informal evaluation of a new syndromic surveillance system using NHS 111 remote health advice data. Methods: NHS 111 syndromic indicators were monitored daily at national and local level. Statistical models were applied to daily data to identify significant exceedances; statistical baselines were developed for each syndrome and area using a multi-level hierarchical mixed effects model. Results: Between November 2013 and October 2014, there were on average 19 095 NHS 111 calls each weekday and 43 084 each weekend day in the PHE dataset. There was a predominance of females using the service (57%); highest percentage of calls received was in the age group 1-4 years (14%). This system was used to monitor respiratory and gastrointestinal infections over the winter of 2013-14, the potential public health impact of severe flooding across parts of southern England and poor air quality episodes across England in April 2014. Conclusions: This new system complements and supplements the existing PHE syndromic surveillance systems and is now integrated into the routine daily processes that form this national syndromic surveillance service.

Internet-based remote health self-checker symptom data as an adjuvant to a national syndromic surveillance system.

Syndromic surveillance is an innovative surveillance tool used to support national surveillance programmes. Recent advances in the use of internet-based health data have demonstrated the potential usefulness of these health data; however, there have been limited studies comparing these innovative health data to existing established syndromic surveillance systems. We conducted a retrospective observational study to assess the usefulness of a national internet-based 'symptom checker' service for use as a syndromic surveillance system. NHS Direct online data were extracted for 1 August 2012 to 1 July 2013; a time-series analysis on the symptom categories self-reported by online users was undertaken and compared to existing telehealth syndromic data. There were 3·37 million online users of the internet-based self-checker compared to 1·43 million callers to the telephone triage health service. There was a good correlation between the online and telephone triage data for a number of syndromic indicators including cold/flu, difficulty breathing and eye problems; however, online data appeared to provide additional early warning over telephone triage health data. This assessment has illustrated some potential benefit of using internet-based symptom-checker data and provides the basis for further investigating how these data can be incorporated into national syndromic surveillance programmes.

Using public health scenarios to predict the utility of a national syndromic surveillance programme during the 2012 London Olympic and Paralympic Games.

During 2012 real-time syndromic surveillance formed a key part of the daily public health surveillance for the London Olympic and Paralympic Games. It was vital that these systems were evaluated prior to the Games; in particular what types and scales of incidents could and could not be detected. Different public health scenarios were created covering a range of potential incidents that the Health Protection Agency would require syndromic surveillance to rapidly detect and monitor. For the scenarios considered it is now possible to determine what is likely to be detectable and how incidents are likely to present using the different syndromic systems. Small localized incidents involving food poisoning are most likely to be detected the next day via emergency department surveillance, while a new strain of influenza is more likely to be detected via GP or telephone helpline surveillance, several weeks after the first seed case is introduced.