Irit Hochberg1, Deeb Daoud2, Naim Shehadeh2,3, Elad Yom-Tov4. 1. Institute of Endocrinology, Diabetes and Metabolism, Rambam Health Care Campus, 8 Ha'Aliya Street, POB 9602, 31096, Haifa, Israel. i_hochberg@rambam.health.gov.il. 2. Institute of Endocrinology, Diabetes and Metabolism, Rambam Health Care Campus, 8 Ha'Aliya Street, POB 9602, 31096, Haifa, Israel. 3. Bruce Rappaport Faculty of Medicine, Technion - Israel Institute of Technology, Haifa, Israel. 4. Microsoft Research, Herzliya, Israel.
Abstract
AIMS: Diabetes is often diagnosed late. This study aimed to assess the possibility for earlier detection of diabetes from search data, using predictive models trained on large-scale data. METHODS: We extracted all English-language queries made by people in the USA to Bing during 1 year and identified queries containing symptoms of diabetes. We compared the ability of four different prediction models (linear regression, logistic regression, decision tree and random forest) to distinguish between users who stated that they were diagnosed with diabetes and users who did not refer to diabetes or diabetes drugs but queried about at least one of the symptoms. RESULTS: We identified 11,050 "new diabetes users" who stated they had been diagnosed with diabetes and approximately 11.5 million "control users" who queried about symptoms without querying for terms related to diabetes. Both the logistic regression and the random forest models were able to distinguish between the populations with an area under curve of 0.92 which translates to a positive predictive value of 56% at a false-positive rate of 1%. The model could identify patients up to 240 days before they mentioned being diagnosed. CONCLUSIONS: Some undiagnosed diabetes patients can be detected accurately according to their symptom queries to a search engine. Such earlier diagnosis, especially in cases of type 1 diabetes, could be clinically meaningful. The ability of search engines to serve as a population-wide screening tool could potentially be improved using additional data provided by users.
AIMS: Diabetes is often diagnosed late. This study aimed to assess the possibility for earlier detection of diabetes from search data, using predictive models trained on large-scale data. METHODS: We extracted all English-language queries made by people in the USA to Bing during 1 year and identified queries containing symptoms of diabetes. We compared the ability of four different prediction models (linear regression, logistic regression, decision tree and random forest) to distinguish between users who stated that they were diagnosed with diabetes and users who did not refer to diabetes or diabetes drugs but queried about at least one of the symptoms. RESULTS: We identified 11,050 "new diabetes users" who stated they had been diagnosed with diabetes and approximately 11.5 million "control users" who queried about symptoms without querying for terms related to diabetes. Both the logistic regression and the random forest models were able to distinguish between the populations with an area under curve of 0.92 which translates to a positive predictive value of 56% at a false-positive rate of 1%. The model could identify patients up to 240 days before they mentioned being diagnosed. CONCLUSIONS: Some undiagnosed diabetespatients can be detected accurately according to their symptom queries to a search engine. Such earlier diagnosis, especially in cases of type 1 diabetes, could be clinically meaningful. The ability of search engines to serve as a population-wide screening tool could potentially be improved using additional data provided by users.
Entities:
Keywords:
Diabetes; Digital health; Internet; Symptoms