Data-driven identification of communities with high levels of tuberculosis infection in the Democratic Republic of Congo

Abstract

When access to diagnosis and treatment of tuberculosis is disrupted by poverty or unequal access to health services, marginalized communities not only endorse the burden of preventable deaths, but also suffer from the dramatic consequences of a disease which impacts one's ability to access education and minimal financial incomes. Unfortunately, these pockets are often left unrecognized in the flow of data collected in national tuberculosis reports, as localized hotspots are diluted in aggregated reports focusing on notified cases. Such system is therefore profoundly inadequate to identify these marginalized groups, which urgently require adapted interventions. We computed an estimated incidence-rate map for the South-Kivu province of the Democratic Republic of Congo, a province of 5.8 million inhabitants, leveraging available data including notified incidence, level of access to health care and exposition to identifiable risk factors. These estimations were validated in a prospective multi-centric study. We could demonstrate that combining different sources of openly-available data allows to precisely identify pockets of the population which endorses the biggest part of the burden of disease. We could precisely identify areas with a predicted annual incidence higher than 1%, a value three times higher than the national estimates. While hosting only 2.5% of the total population, we estimated that these areas were responsible for 23.5% of the actual tuberculosis cases of the province. The bacteriological results obtained from systematic screenings strongly correlated with the estimated incidence (r = 0.86), and much less with the incidence reported by epidemiological reports (r = 0.77), highlighting the inadequacy of these reports when used alone to guide disease control programs.

Figure 1

Mapping of TB predicted incidence rate for the South-Kivu province of the Democratic Republic of Congo (inlay). No color: below 0.1%; Yellow : between 0.1% and national average (0.322%); Orange : between 0.322% and 1%; Red : above 1%. Geospatial data provided by OpenStreetMap (under ODbL license) and Natural Earth (public domain), map produced using Cartopy v0.20 from https://scitools.org.uk/cartopy.

Figure 2

The Mediscout mobile and web applications. (A) Mobile application with interface for mission acceptance by the community health-worker. (B) Questionnaire in the mobile application. (C) Statistics of the missions in the web application (note the mapping of the screenings at the bottom).

Figure 3

Distribution of individual risk for the population located in low risk zones (blue) where the predicted incidence rate is lower than 1% and for the population in high risk zones (orange). Note that the proportion of lower scores ($<4$) is higher in the lower risk communities (bottom). Score medians in both subpopulations (3 and 5 respectively) with quartiles are reported in the upper box plot.

Figure 4

Proportion on laboratory confirmed TB cases per score class.

Figure 5

Correlation of the predicted incidence with the measured incidence within sampled population (right). Note the high value of the correlation coefficient r. The parameter a represent the slope of the fitting line. As a term of comparison, the correlation of the incidence extracted from the health system reports with the measured incidence within the sampled population (left). Note in this case the lower correlation coefficient and the slope of the fitting line (the measured incidence surpass 10x the incidence of reported cases).

Figure 6

Expected sensitivity and specificity if the threshold to refer subjects to lab is changed to a different value. The vertical lines correspond to a threshold of 4 and 6. A threshold of 4 is maybe too conservative; in other contests, one can safely use 6.