[Paper Review] Clinically verified pre-screening for cancer using web search queries: Initial results.

Elad Yom‐Tov|arXiv (Cornell University)|Feb 21, 2018

Data-Driven Disease Surveillance3 citations

TL;DR

This study proposes using web search queries from individuals who self-identify as potentially having lung, breast, or colon cancer—recruited via Bing ads—to predict cancer risk. By correlating past queries with clinically verified risk scores, an automated classifier achieved an AUC of 0.64 for all cancers and 0.76 for colon cancer, demonstrating search queries as a viable tool for early cancer screening.

ABSTRACT

Search engine queries have been demonstrated to be a useful signal for screening people for different cancer types. Past work focused on a biased population which indicated that they were suffering from the condition, or else inferred which people had the condition of interest using their queries. Here we used a combination of an online advertising campaign and a clinically verified questionnaire to identify at-risk people, and correlated their past queries with these risk scores. People who suspected they were suffering from lung, breast, or colon cancer were recruited through ads shown on the Bing search engine to complete a clinically verified risk questionnaire. Of those, 201 people agreed to participate in the research and their past queries could be obtained. An automated classifier to predict their risk score based on past queries reached an Area Under the ROC (AUC) of 0.64 for all cancers, and 0.76 for colon cancer. These results demonstrate the utility of search engine queries to screen for cancer and are the represent the first step in utilizing advertising systems to screen for cancer and detect it earlier than has been previously possible.

Motivation & Objective

• To develop a method for identifying individuals at risk of cancer using their past web search queries.
• To recruit participants who suspect they may have lung, breast, or colon cancer through targeted online advertising.
• To clinically verify self-reported risk using a standardized questionnaire to ensure data quality.
• To evaluate the predictive power of search queries for cancer risk using machine learning.
• To demonstrate the feasibility of using advertising platforms to enable early cancer detection through search behavior.

Proposed method

• Recruited participants via targeted Bing search engine advertisements shown to users who may be at risk of cancer.
• Collected self-reported risk scores using a clinically verified questionnaire administered online.
• Accessed and analyzed participants' anonymized past search queries from their Bing search history.
• Trained an automated classifier to predict cancer risk based on the content and patterns in past search queries.
• Evaluated model performance using the Area Under the ROC Curve (AUC) for overall cancer and specific cancer types.
• Used a representative sample of 201 participants who consented to share their search history and complete the questionnaire.

Experimental results

Research questions

• RQ1 Can search engine queries reliably predict self-reported cancer risk among individuals who suspect they may have cancer?
• RQ2 How effective is a machine learning classifier in predicting cancer risk using anonymized past search queries?
• RQ3 What is the predictive performance of search queries for specific cancer types, particularly colon cancer?
• RQ4 Can online advertising platforms be effectively used to recruit at-risk individuals for cancer screening research?
• RQ5 How does the performance of query-based screening compare to traditional methods in early detection?

Key findings

• The automated classifier achieved an Area Under the ROC Curve (AUC) of 0.64 when predicting overall cancer risk across all three cancer types.
• For colon cancer specifically, the classifier achieved an AUC of 0.76, indicating strong predictive performance.
• The study successfully recruited 201 participants through a Bing advertising campaign who consented to share their past search queries.
• Clinically verified risk scores were obtained via a standardized questionnaire, ensuring data reliability.
• This is the first study to use an advertising platform to recruit at-risk individuals and validate search queries for cancer screening.
• The results demonstrate that search queries can serve as a useful, non-invasive signal for early cancer detection.