Impact of human mobility on the emergence of dengue epidemics in Pakistan

Abstract

The recent emergence of dengue viruses into new susceptible human populations throughout Asia and the Middle East, driven in part by human travel on both local and global scales, represents a significant global health risk, particularly in areas with changing climatic suitability for the mosquito vector. In Pakistan, dengue has been endemic for decades in the southern port city of Karachi, but large epidemics in the northeast have emerged only since 2011. Pakistan is therefore representative of many countries on the verge of countrywide endemic dengue transmission, where prevention, surveillance, and preparedness are key priorities in previously dengue-free regions. We analyze spatially explicit dengue case data from a large outbreak in Pakistan in 2013 and compare the dynamics of the epidemic to an epidemiological model of dengue virus transmission based on climate and mobility data from ∼40 million mobile phone subscribers. We find that mobile phone-based mobility estimates predict the geographic spread and timing of epidemics in both recently epidemic and emerging locations. We combine transmission suitability maps with estimates of seasonal dengue virus importation to generate fine-scale dynamic risk maps with direct application to dengue containment and epidemic preparedness.

Keywords: Pakistan; dengue; epidemiology; human mobility; mobile phones.

Fig. 1.

Human mobility dynamics in Pakistan. (A) Population density (red, high density; yellow, low density) and mobile phone tower coverage from the mobile phone operator in Pakistan (colored in gray) per tehsil. (B) The top routes of travel between pairs of tehsils in Pakistan. A line is drawn if at least 20,000 trips occurred between the origin and destination between June and December 2013. The top routes occur between Karachi and cities in northern Punjab province, particularly Lahore tehsil. (C) Relative direction and volume of travel. For each trip, we calculated the distance traveled from the origin and the destination. The origin location was centered at 0,0 and the longitude distance and latitude distance to the destination are shown. Although many trips occurred over short distances, a substantial amount of travel occurred between the southeastern and northern parts of the country, reflecting the geography and population distribution of Pakistan.

Fig. S1.

The mobility values derived from the mobile phone data. (A) The mobility estimates for travel from Karachi to Lahore. Using the mobile phone data, we calculated both the number of trips between tehsils (A, Top Left) and the daily number of active subscribers: flux value (A, Top Right). The number of trips was normalized by the daily number of active subscribers (A, Bottom Left) that was used to estimate the importation of infected travelers from Karachi. Our dataset runs from June to December 2013 and we inferred the mobility estimates (gray) for January–June of that year (Materials and Methods) (A, Bottom Right). (B) The most frequently traveled routes based on the mobile phone data. We compared the most traveled routes (top 10,000 routes) 1 mo before Ramadan (B, Left) (June), during Ramadan (July) (B, Center), and after Ramadan (B, Right) (September). The top routes remain consistent throughout the major holiday, in particular the large amount of travel between Karachi and Punjab province. A line is drawn if more than 10,000 trips were taken over a route. (C) Country-wide population flux values per day. Total population flux per day was measured (red line). Ramadan is shown in gray. We have adjusted the mobile phone data (CDRs) (black line shows adjustment) to account for a sharp drop in the raw CDRs after day 145 and fitted the adjusted data [spline, blue line; moving average (MA) trend, purple line], using two common time series methods (a MA and smoothing spline) to smooth the data. For the remaining analysis, we use these trend lines from these time series smoothing methods.

Fig. 2.

Dengue epidemiology in Pakistan in 2013. (A) The location of dengue cases throughout Pakistan and the number of cases per week by tehsil. Tehsils that reported at least 15 cases are shown on the map with corresponding color shown in the time series. The majority of cases were reported in Karachi (gray), Lahore (blue), and Mingora (orange). The dengue season in the entire country lasted 35 wk, with the first reported case in Karachi during week 18 (end of April). (B) The reported cases (red), temperature (blue), and model fit (black) for Karachi are shown. Using the case and temperature data, the human and vector population dynamics were modeled (Materials and Methods).

Fig. 3.

Mobility estimates derived from mobile phone data predict the timing of introduced cases around the country that spark epidemics. (A and B) The estimated introduced cases from Karachi to (A) Lahore (total dengue cases: 1,538) and (B) Mingora (total dengue cases: 4,029). The estimated introductions (assuming 30% of individuals travel, a 2% reporting rate, and a probability of 0.01) from the mobile phone data (boxplot in blue), the diffusion model (boxplot in green), actual case data (red), and estimated dengue suitability (gray) are shown. Dengue suitability was defined based on temperature and relative humidity, using a measure that is linearly proportional to vectoral capacity (Materials and Methods). Values near zero are unsuitable for dengue transmission. For Lahore and Mingora, the estimated introduction from the case data alone is shown (red cross-hatched box). In all instances, the mobile phone data were able to predict the timing of the first introduced case in each tehsil. An arrow indicates the week of the first reported case in each tehsil.

Fig. S2.

The various distance measures and gravity model fits to the mobile phone data. We calculated the Euclidean distance (kilometers), road distance (gray, in kilometers), and travel time distance (red, in minutes) between tehsil centroids. (A) The three measures were highly correlated (Pearson’s correlation coefficient between Euclidean distance and road distance, 0.988; between Euclidean distance and travel time distance, 0.99). We then fitted gravity models to the mobile phone data, using each distance measure: (B) Euclidean distance, (C) road distance, and (D) travel time distance. Shown is the normalized predicted amount of travel from each gravity model compared to the mobile phone data.

Fig. 4.

The estimated spatial spread of imported dengue. Using the modeled dengue dynamics in Karachi and mobility measured from the mobile phone data or a diffusion model, we estimated the time of the first introduced case to the rest of the country. The mobile phone data predict the earliest introductions in eastern Pakistan near Lahore and inland toward Swat Valley (Mingora). In comparison, the mobility model predicts early introductions in southern Pakistan with few introductions in Mingora. These differences are highlighted in the difference in predictions plot—the number of days earlier (red) from the mobile phone predictions or earlier (yellow) from the diffusion model (without the mobile phone data).

Fig. 5.

Dynamic risk mapping for dengue epidemics in Pakistan. (A) Average dengue suitability in Pakistan during 2013 based on temperature and humidity. The southern part of the country is the most suitable for dengue (high values are shown in red). (B) Epidemic risk, measured as a composite of vector suitability and frequency of introductions from endemic areas in southern Pakistan. Places in red are the most at risk whereas those in blue have low risk of epidemics. (C) Variation in the timing of dengue risk by location (per tehsil). A box is drawn if suitability multiplied by estimated introduction events is greater than 0, indicating a nonnegative risk. Each line shown is a tehsil, with the y axis corresponding to the ranked distance from Karachi. The dengue risk is shown using the mobile phone data (blue) or the diffusion model (green) to estimate introduction events from Karachi. In general, the mobile phone data predict earlier risk to tehsils farther away from Karachi, in particular tehsils in Punjab (Faisalabad and Lahore, for example), than the diffusion model data. The diffusion model predicts the earliest risk to nearby tehsils such as Malir.

Fig. S3.

The temperature data from weather stations in Pakistan. (A) The location of weather stations in Pakistan. (B) Daily temperature (in Celsius) from selected weather stations in Pakistan. (C) The temperature-dependent parameters used in the ento-epidemiological framework for Karachi.

Fig. S4.

Sensitivity analysis on the carrying capacity (K) fitted to the biting rate (a) for each value of K, using a maximum-likelihood framework. The case data for Karachi are shown as points.

Fig. S5.

Sensitivity analysis of estimated introductions to tehsils based on the CDR varying the percentage that travel or the probability of introduction. We analyzed the estimated introductions, using the mobile phone data to quantify travel from Karachi to other tehsils. We varied the percentage of travel from Karachi (Left) and the probability of introduction (Right).

Fig. S6.

Mobility models derived from mobile phone data predict the timing of introduced cases around the country that spark epidemics. Shown are the estimated introduced cases from Karachi to all tehsils that have reported at least 15 cases. The estimated introductions from the mobile phone data (blue), the diffusion model (orange), the gravity model (green), and actual case data (red) are shown. Introductions are estimated assuming that 30% of individuals travel consistent with the mobile phone data (prob = 0.01). If there were no estimated introductions from both the mobile phone data and the diffusion model, no boxes are drawn. In the majority of the KPK tehsils, both the mobile phone data and the diffusion model did not predict any introductions, likely because these cases were associated with the initial outbreak in Mingora (Alai–Peshawar). However, in the Punjab tehsils (Faisalabad, Ferozawala, Rawalpindi, and Sheikhupura) the mobile phone data consistently predicted an earlier introduction than the diffusion model and before the date of the first reported case.

Fig. S7.

The estimated introductions for Mingora/Babuzai tehsil and Lahore while varying parameters. (A and B) The estimated introductions for Mingora/Babuzai and Lahore assuming various reporting rates for Karachi. Shown are the estimated introduced cases from Karachi to Mingora and Lahore (assuming 30% of individuals travel and a probability of 0.01) with various reporting rates: 2%, 3%, 5%, and 10%. Introduced cases were estimated using the mobile phone data (blue), a gravity model fitted to the mobile phone data (orange), or a diffusion model (green). The mobile phone data consistently predict earlier introductions than either the gravity model or the diffusion model, with the diffusion model predicting the latest introductions of the three methods. In particular, the diffusion model does not report introductions into Mingora and for high reporting rates (10%), there are no predicted introductions into this tehsil. (C and D) The estimated introductions for Mingora/Babuzai and Lahore, assuming various percentages of individuals travel from Karachi. Shown are the estimated introduced cases from Karachi to Mingora and Lahore (assuming a 2% reporting rate and a probability of 0.01) with various percentages of individuals traveling from Karachi: 10%, 20%, and 30%. Unsurprisingly, as the percentage of individuals who travel increases, the number of importations to both Mingora and Lahore increases. The timing of these introductions also increases as the percentage of individuals who travel from Karachi increases. (E and F) The estimated introductions for Mingora and Lahore, assuming various probabilities. Shown are the estimated introduced cases from Karachi to Mingora and Lahore (assuming 30% of individuals travel and a 2% reporting rate) with various probabilities: 0.001, 0.01, 0.1, and 0.9 (Materials and Methods).

Fig. S8.

The estimated introductions to Lahore. Using the mobile phone data, the number of introductions from (A) Karachi and (B) Mingora to Lahore are shown. As in Fig. 3, we estimated the introduced cases from Karachi or Mingora (Swat) to Lahore. (C) A comparison of the median number of introductions per week.