[click here and type title]

International Biometric Society

Nowcasting daily mortality on real-time for public health surveillance
Liliana Antunes^1,2,3, Marília Antunes², Baltazar Nunes^1,3
1 Department of Epidemiology, Instituto Nacional de Saúde Dr. Ricardo Jorge, Portugal, 2 Centre of Statistics and Applications of the University of Lisbon ^*, Portugal 3 FCT Research Unit of the Institute of Preventive Medicine of the University of Lisbon, Portugal
It is well described in the literature that during health-related events, like influenza epidemics or occurrence of extreme weather events, such as heat waves or cold spells, the observed “all causes” mortality increases above the expected, presenting excess deaths of 20-40%.These facts have triggered the development of Public Health surveillance systems, with the objective of early detection of such events, so that public health measures can be taken in time, as well the estimation of its impact. Since the 2003 heat wave, the Department of Epidemiology of INSA has been developing a daily mortality monitoring system (VDM), which came fully operational in 2007. In this system INSA receives electronically, on a daily basis, all deaths registered in the previous day at the Conservatories of the Civil Registry. Previous studies have showed that 95% of the deaths that occurred in specific day are registered up to 6 days after, with an average and median delay of 8 and 2 days respectively. Given that the detection of the impact of the events are dependent on these delays, it would be very important to have, on a daily basis, the estimate of the number of deaths occurred but not yet registered in the system.

In this context our objective was to develop a statistical model to predict (nowcast) on a daily basis, the number of deaths that occurred on day t, using the registered deaths up to the current day, t+i, i=1,2,… . The developed method updates the expected number of deaths on day t by modelling the probability that a death on that day t is reported until day t+i , being i the reporting delay in days, measured as the difference between the date of death and the date of reporting to INSA. The real number of deaths on day t, M_t, is estimated by M_t*=n_i/p_i, where n_i is the number of deaths on day t reported to the system until day t+i, and p_i is the probability that a death occurred at day t is reported until day t+i. Logistic regression models were used to estimate each p_i using the weekday of death and the binary variables occurrence of a public holiday and occurrence of health-related event (influenza epidemic or extreme weather event). The relative error associated with each day of prediction was estimated by |M_t-M_t*|/M_t. Models were adjusted using mortality data collected from September 2008 to September 2012, comprising a total of 454,422 observations.

Applying the calibrated model to the period from October 2012 to August 2013, 55% and 94% of predictions were within a 10% error interval of the real number of deaths, on the first (t+1) and second (t+2) days of prediction, respectively. The median error was 9.4%, 3.4% and 2.5%, on the first (t+1), second (t+2) and third (t+3) days of prediction, respectively. During the 2013 heat wave it was possible to detect the first day of excess mortality on prediction day t+1. The heat wave mortality peak was observed on July 8^th with 500 deaths. After two days (t+2), 360 deaths have been reported. Our method predicted 472 deaths, corresponding to a prediction error of 5.6%.

Therefore, our results showed that the application of this method reduces substantially the time between the onset of health-related event and its detection (timeliness) by the daily mortality surveillance system, allowing the early detection and estimation of the impact of health-related events like heat waves.

^* This work was financially supported by national funds from FCT: (2) projects PEst-OE/MAT/UI 0006/2011 and PTDC/MAT/118335/2010

International Biometric Conference, Florence, ITALY, 6 – 11 July 2014