Algorithms for verbal autopsies: a validation study in Kenyan children.

Introduction Accurate information on the causes of mortality is necessary for effective planning and evaluation of health care programmes (1-3). In recent years, the verbal autopsy (VA) questionnaire has been widely used for collecting such information in situations where the medical certification of deaths in childhood is incomplete. Trained fieldworkers interview bereaved relatives using a structured questionnaire in order to elicit information on the symptoms their child experienced before death. The information from completed questionnaires is then summarized and interpreted to give a likely cause of death for each child. Probably the most common method for ascribing causes of death from VA questionnaires is when the completed questionnaires are reviewed by one or more physicians who ascribe probable causes of death (4-10). All parts of the questionnaire, particularly any open-ended sections, are thus incorporated into the diagnosis. However, large-scale surveys can prohibit the use of long, detailed questionnaires; furthermore, open-ended questions, asked by lay interviewers, may prove difficult in establishing case histories. An alternative means of ascribing causes of death is to follow a set of pre-defined diagnostic criteria in an expert algorithm (11-14). In this case, the questionnaire consists of closed questions and yields only pre-coded information. Moreover, as the algorithm uses well-defined diagnostic criteria to ascribe causes of death, changes in cause-specific mortality may be compared over time or between different studies (2). Expert algorithms have been the subject of validation studies in the Philippines (170 deaths) (11) and in Namibia (135 deaths) (13). The accuracy of these algorithms was estimated by comparing the ascribed causes of death with the medically confirmed diagnoses of children who died in hospital. For deaths due to measles and malnutrition, which have signs and symptoms that are readily recognized by lay persons, the algorithms gave a relatively high sensitivity and specificity (Table 1). However, deaths due to other causes, most notably malaria and acute respiratory infection (ARI), were not assigned accurately in either study population. [TABULAR DATA 1 OMITTED] Poor diagnostic performance of the VA technique may be due to shortcomings of the questionnaire or of the method used to ascribe the cause of death. Information on symptoms leading to a child's death may not be elicited because the symptoms are not easily recognized by lay persons, or are poorly recalled by relatives, or are not included in the questionnaire (1, 15). Furthermore, it may not be possible to discriminate between certain causes of death on the basis of signs and symptoms alone (16, 17). The VA technique may be applied with high diagnostic accuracy in situations where each cause of death is always preceded by a unique set of signs and symptoms (2). This set of signs and symptoms may be found using standard statistical methods. For given data, an algorithm is then obtained which results in the highest possible sensitivity and specificity. These techniques have been used in diagnostic methods and screening (18,19). Clearly, algorithms can only be derived when the medically confirmed causes of death are available, although they may then be used more widely. We report here on algorithms derived using logistic regression and compare their diagnostic accuracy with that from several expert algorithms. Methods Study population and prospective surveillance of hospital deaths A prospective VA study was conducted between May 1989 and April 1993 at Kilifi District Hospital, 60 km north of Mombasa on the Kenyan coast. Details of the study population and the VA methods have been described elsewhere (8). In brief, all children admitted to the paediatric ward at Kilifi District Hospital were examined on admission and during their stay in hospital. Full clinical examinations and laboratory investigations were carried out on each child and recorded on a standard proforma. …