The First Predictor of Impending Stroke: An Internet Search-Based Algorithm Blind to Established Cardiometabolic Risk, Outperforms Classical Risk Factor-Based Calculators

Cerebrovascular disease is a leading cause of mortality and disability and an immense global burden. It is partly related to aging in a metabolic syndrome-promoting environment. Prevention strategies are insufficient: they rely on intermittent screening in predominantly high-risk individuals, while most cases occur in the intermediate risk population. Screening algorithms like the Framingham Risk Score predict events in the next decade using traditional risk factors (age, diabetes, hypertension, obesity, hyperlipidemia, atrial fibrillation and smoking). No current tool can predict a near/impending stroke in the next few weeks/months. This could provide a time window for urgent preventive measures (anticoagulation, hypertension control or lipid lowering). Algorithms analyzing daily computer usage, examining motor, language and executive functions, correlate with standard cognitive tests. Covert cerebrovascular disease is linked to subtle cognitive and motor deficits and increased risk for stroke. We examined whether this could be harnessed to predict an imminent stroke. We developed an algorithm based on internet search queries to identify people at increased risk for a near stroke event. The algorithm was entirely blind to traditional cardiovascular risk factors or risks for cognitive decline. We analyzed queries submitted to the Bing search engine by 285 people for which a stroke event was identified and 1195 controls, with regards to attributes previously shown to reflect cognitive function. Controls included random people 60 years and above or similar aged individuals who queried for one of eight control conditions: myocardial infarction, atrial fibrillation, hypertension, migraine, B12 deficiency, depression, hypothyroidism and surgery. We used a random forest model with 1000 trees to distinguish the patient cohort from each of the control cohorts. Ten-fold cross-validation at the user level was used to reduce the likelihood of overfitting. All processing was conducted using Matlab 2019. Results: Our model performed well against all controls with area under the curve of the receiver operating curve (AUC) of 0.985 or higher and a true positive rate (at 1% false positive rate) above 80% for separating patients from each of the controls. The predictive power rose as the stroke date approached and if data was acquired beginning 120 days prior to the event. Good prediction accuracy was also obtained for a prospective cohort of users collected one year later. We propose that impending stroke can be identified through alterations in internet usage patterns that reflect cognitive function, with high predictive power (AUC exceeding 0.985 for a near event in comparison to an AUC of ~0.71 in the Framingham Stroke Risk Score for a 10 year event). This could potentially be applied inexpensively, continuously and on a large scale to willing individuals with the aim of reducing stroke events.