Developing Tumor Diagnosis Models With Verbal Autopsy

Background: The verbal autopsy (VA) is used to collect information on cause of death (COD) from bereaved relatives. A COD may be diagnosed by expert assessment, expert algorithm and data-derived algorithm. In rural China, there were 80% people died at home, so the validity of COD needed to be estimated.Objectives: This paper developed Bayes probabilistic diagnosis models for interpreting verbal autopsy data to infer the COD of patients suffered from tumors, especially some common malignant tumors, including carcinoma of esophagus, stomach, colon, rectum, anus, liver and bronchus. We could infer the accuracy of expert reviews and death certificate in rural China by comparing with the diagnostic results from these diagnosis models. And also we could infer COD of rural China with diagnosis models.Methods: We considered the death proportion of Disease Surveillance Point (DSP) of China in 2000 as prior probability, which was adjusted by the under-reporting rate. The probabilities of symptoms of each disease diagnosed by gold standard were defined as the conditional probabilities. The posterior probability can be deduced using Bayes's formula, thus we could develop the diagnosis models of tumors (uncategorized) and some certain common malignant tumors that we focused on.Results:â—†The diagnostic accuracy of Bayes probabilistic model for inferring the cause of death of tumors (uncategorized) was higher than the other models, and this model was stable. The accuracy of oesophagus carcinoma modle was high, but the model was not stable. As to the models for carcinoma of stomach, liver and bronchus, the diagnostic accuracy was poor. And the diagnosis model for colon, rectum and anus couldn't be developed. â—†The diagnostic accuracy of whole-question VA expert review was very high, but the accuracy of structured-question VA expert review was not high, and was equal to that of the models.â—†The specificity and consistency of the death certificates in rural places of China was very high, but the sensitivity was very low.Conclusions:â—†We could use the diagnosis model for tumor (uncategorized) to infer the COD of every VA questionnaire, but the specialized carcinoma models seem not appropriate to give diagnoses.â—†The diagnosis accuracy of expert reviews is very high, but it is easy to be affected by open questions.â—†In rural places of China, the diagnosis models of tumors are valuable to infer the COD of tumor sufferers.