Page 29 - Contributed Paper Session (CPS) - Volume 3

CPS1934 Atikur R. K. et al.
apply a model that incorporates a time lag relationship between response and
predictor variables.
Diurnal temperature change (difference between the maximum and
minimum daily temperature) is a measure for variation in temperature change
and has a greater impact on respiratory tract infection (RTI) episodes. Changes
in temperature, precipitation, relative humidity, and air pollution influence viral
activity and transmission and may contribute to the size and severity of the
epidemics (Eccles, 2002; Mirsaeidi et al., 2016; Zhang et al., 2000). A good
forecast on the severity of the epidemic is likely to provide good information
for healthcare planning at least for at-risk group. To forecast weekly number
of hospitalization in São Paulo city, Alencar (2018) applied generalized additive
model with autoregressive terms (GAMAR) and Poisson distribution and
demonstrated that inclusion of seasonal parameter in the model provides
better forecasting performance than GAMAR with binomial distribution.
However, these models do not consider time lag relationship between the
response and predictor variables constructed from climatic factors.
Furthermore, our dataset is in panel form with daily number of RTI episodes
for 35 weather stations. Thus, we consider panel generalized linear models and
machine learning methods for prediction of daily RTI episodes.
To organize the rest of our paper, we discuss the data integration
procedure and computation for rolling time series statistics of climatic factors
in Section 2. In Section 3, we provide our analytic results to demonstrate the
effect of rolling time series statistics on RTI episodes by fitting panel
generalized linear model and provide 7-day ahead forecast along with 1-day
ahead confirmation forecast from regression tree and random forest models.
In Section 4, we discuss on our overall findings and provide concluding
remarks.

2. Methodology
Clinical data and climatic data are integrated to form data tables.
Prescriptions collected from different clinics by 4P Ltd., a prescription audit
and marketing company, over two years have been used in this study. Number
of daily RTI episodes counted from these prescriptions is then integrated with
climatic data obtained from 35 weather stations across the whole country.
Daily episodes of diseases of a particular area are linked with the daily climatic
time series from the weather station closest to that area. Our data integration
procedure has been presented in Figure 1.

18 | I S I W S C 2 0 1 9