Deaths ascribed to non-communicable diseases among rural kenyan adults are proportionately increasing: Evidence from a health and demographic surveillance system, 2003-2010

Background: Non-communicable diseases (NCDs) result in more deaths globally than other causes. Monitoring systems require strengthening to attribute the NCD burden and deaths in low and middle-income countries (LMICs). Data from health and demographic surveillance systems (HDSS) can contribute towards this goal. Methods and Findings: Between 2003 and 2010, 15,228 deaths in adults aged 15 years (y) and older were identified retrospectively using the HDSS census and verbal autopsy in rural western Kenya, attributed into broad categories using InterVA-4 computer algorithms; 37% were ascribed to NCDs, 60% to communicable diseases (CDs), 3% to injuries, and ,1% maternal causes. Median age at death for NCDs was 66y and 71y for females and males, respectively, with 43% (39% male, 48% female) of NCD deaths occurring prematurely among adults aged below 65y. NCD deaths were mainly attributed to cancers (35%) and cardiovascular diseases (CVDs; 29%). The proportionate mortality from NCDs rose from 35% in 2003 to 45% in 2010 (x2 linear trend 93.4; p,0.001). While overall annual mortality rates (MRs) for NCDs fell, cancer-specific MRs rose from 200 to 262 per 100,000 population, mainly due to increasing deaths in adults aged 65y and older, and to respiratory neoplasms in all age groups. The substantial fall in CD MRs resulted in similar MRs for CDs and NCDs among all adult females by 2010. NCD MRs for adults aged 15y to ,65y fell from 409 to 183 per 100,000 among females and from 517 to 283 per 100,000 population among males. NCD MRs were higher among males than females aged both below, and at or above, 65y. The HDSS study site is located in a rural part of Siaya County, in western Kenya [20], [27]. The area consists of 385 villages spread over a 700 km2 area along the shores of Lake Victoria, with a population in 2010 of 224,500 adults in aged 15 years (y) and above. The population, mainly subsistence farmers, are almost exclusively members of the Luo ethnic group and traditionally polygynous, have been described in detail elsewhere [20], [28]. The HDSS is a population-based system with GPS locational data, that longitudinally records demographic (births, deaths, pregnancies, and in- and out-migrations) information [20]. Household census among the population takes place tri-annually, in January-March, May-August, and October-December, by field staff who visit all households in the study site. Verbal autopsy is conducted in subpopulations covered by HDSS. All deaths in residents, defined as having resided in the area for at least four consecutive months, are identified by local village reporters through ongoing local monitoring, and are validated during the tri-annual census. At least one month following death, and within four months to reduce recall bias, an interviewer returns to the home and records events surrounding the death, using standardized WHO verbal autopsy (VA) questionnaires endorsed by the INDEPTH Network [20], [29], [30], with spouses or another close relative of the deceased. Resident identification numbers allow linkage of each death with HDSS data. In this paper, cause of death was attributed using the InterVA-4 methodology, a new public-domain probabilistic model for interpreting cause of death from VA data [31], [32]. This methodology attributes cause of death compatible with the International Classification of Diseases 10 (ICD-10) categorised into 62 overall groups through a computer simulated algorithm. ICD-10 codes and the respective VA coding and disease categories are listed in Table S1. Indicators required to run the InterVA-4 Model were extracted from VA data and entered into the model to generate cause of death. The model produces a maximum of three probable causes of death and their corresponding likelihoods. In this paper, analyses focus on primary cause of deaths since only 10% received a secondary and <1% received a tertiary diagnosis. The model has a built-in facility to adjust for the prevalence of malaria and HIV/AIDS. Before running, the model was set high for both the diseases. Previous studies in our surveillance area have reported the prevalence of malaria and HIV/AIDs at 33% and 14% respectively [20], [21], [25]. The HDSS protocol and consent procedures, including surveillance and VA, are approved by the Ethical Review Committee of the Kenyan Medical Research Institute (#1801) and by the Centers for Disease Control and Prevention Institutional Review Board (#3308). Following cultural customs, compound heads provide written informed consent for all compound members to participate in HDSS activities. Individuals can refuse to participate at any time. All HDSS census and VA data are maintained on a secure server accessed by authorized researchers only. Named data are securely stored in a MS-SQL database and only authorized data personnel have access rights. Datasets analysed by scientists are stripped of names to protect identity. For this evaluation, adults were defined as persons aged 15 years and older. Data were extracted from the HDSS database for all adult deaths in residents, generated from the adult VA questionnaire, between January 2003 and December 2010. Primary cause was derived from aggregated ICD-10 codes generated by the InterVA-4 algorithms [31] (Table S1). Median age of death is presented with interquartile ranges (IQR), for grouped causes of NCD deaths. Analyses are stratified by sex, and into two age groups, using 65 years as the break-point (i.e. below; 15y to <65y, and at or above; ≥65y), to investigate trends in the causes and proportion of premature (aged below 65y) NCD deaths. Descriptive data include deaths in Karemo villages captured 2008–2010, but time trends on the absolute number of deaths and mortality rates 2003–2010 exclude these villages. Mortality rates per 100,000 population for CDs and NCDs, and for main NCD causes (as aggregated through ICD-10 codes; Table S1), were estimated by year and age category using mid-year population-point estimates generated from the HDSS census. The age-sex structure of the adult population was examined per year to clarify if relative proportions changed over time. The highly stable population profile precluded the need to make temporal adjustments to the denominator for analysis of rates. Key social and demographic characteristics generated from questioning the compound head during HDSS census surveys were examined, to compare differences among deaths from NCD and CD, and by sex. This included marital status (ever married, married at time of death, divorced), education (completed primary, secondary school), and socio-economic status (SES). SES quintiles were based on multiple correspondence analysis (MCA) generated from biennial surveys on wealth indicators, reported elsewhere [20], [26]. In this paper we collapsed the five SES quintiles into two, portraying poorest (lowest two quintiles) and less poor (highest three quintiles). Analyses were conducted using SPSS for Windows (Release v21.0; IBM, Endicott, NY, USA), and EpiInfo Stat Calc (v7; CDC Atlanta, USA). Chi-squared (χ2) test for linear trend (LT) determined the significance of changing rates by sex and age over time (2003 to 2010). Pearson's χ2 test was used to determine differences between groups. Mantel Haenszel Relative Risks (RR), with Taylor Series 95% confidence intervals (CI), were used to compare annual mortality rates between sexes. We stratified RR analyses for mortality rates by age groups, sex, and year generating a summary χ2, with a Mantel Haenszel weighted RR (MHRR) and Greenlands-Robins 95% CI. Significance was set at 5% or less.