RESUMEN
BACKGROUND: Depression and anxiety disorders among the global population have worsened during the COVID-19 pandemic. Yet, current methods for screening these two issues rely on in-person interviews, which can be expensive, time-consuming, and blocked by social stigma and quarantines. Meanwhile, how individuals engage with online platforms such as Google Search and YouTube has undergone drastic shifts due to COVID-19 and subsequent lockdowns. Such ubiquitous daily behaviors on online platforms have the potential to capture and correlate with clinically alarming deteriorations in depression and anxiety profiles of users in a noninvasive manner. OBJECTIVE: The goal of this study is to examine, among college students in the United States, the relationships of deteriorating depression and anxiety conditions with the changes in user behaviors when engaging with Google Search and YouTube during COVID-19. METHODS: This study recruited a cohort of undergraduate students (N=49) from a US college campus during January 2020 (prior to the pandemic) and measured the anxiety and depression levels of each participant. The anxiety level was assessed via the General Anxiety Disorder-7 (GAD-7). The depression level was assessed via the Patient Health Questionnaire-9 (PHQ-9). This study followed up with the same cohort during May 2020 (during the pandemic), and the anxiety and depression levels were assessed again. The longitudinal Google Search and YouTube history data of all participants were anonymized and collected. From individual-level Google Search and YouTube histories, we developed 5 features that can quantify shifts in online behaviors during the pandemic. We then assessed the correlations of deteriorating depression and anxiety profiles with each of these features. We finally demonstrated the feasibility of using the proposed features to build predictive machine learning models. RESULTS: Of the 49 participants, 49% (n=24) of them reported an increase in the PHQ-9 depression scores; 53% (n=26) of them reported an increase in the GAD-7 anxiety scores. The results showed that a number of online behavior features were significantly correlated with deteriorations in the PHQ-9 scores (r ranging between -0.37 and 0.75, all P values less than or equal to .03) and the GAD-7 scores (r ranging between -0.47 and 0.74, all P values less than or equal to .03). Simple machine learning models were shown to be useful in predicting the change in anxiety and depression scores (mean squared error ranging between 2.37 and 4.22, R2 ranging between 0.68 and 0.84) with the proposed features. CONCLUSIONS: The results suggested that deteriorating depression and anxiety conditions have strong correlations with behavioral changes in Google Search and YouTube use during the COVID-19 pandemic. Though further studies are required, our results demonstrate the feasibility of using pervasive online data to establish noninvasive surveillance systems for mental health conditions that bypasses many disadvantages of existing screening methods.
RESUMEN
BACKGROUND: The identification of surgical site infections for infection surveillance in hospitals depends on the manual abstraction of medical records and, for research purposes, depends mainly on the use of administrative or claims data. The objective of this study was to determine whether automating the abstraction process with natural language processing (NLP)-based models that analyze the free-text notes of the medical record can identify surgical site infections with predictive abilities that match the manual abstraction process and that surpass surgical site infection identification from administrative data. METHODS: We used surgical site infection surveillance data compiled by the infection prevention team to identify surgical site infections among patients undergoing orthopaedic surgical procedures at a tertiary care academic medical center from 2011 to 2017. We compiled a list of keywords suggestive of surgical site infections, and we used NLP to identify occurrences of these keywords and their grammatical variants in the free-text notes of the medical record. The key outcome was a binary indicator of whether a surgical site infection occurred. We estimated 7 incremental multivariable logistic regression models using a combination of administrative and NLP-derived variables. We split the analytic cohort into training (80%) and testing data sets (20%), and we used a tenfold cross-validation approach. The main analytic cohort included 172 surgical site infection cases and 200 controls that were repeatedly and randomly selected from a pool of 1,407 controls. RESULTS: For Model 1 (variables from administrative data only), the sensitivity was 68% and the positive predictive value was 70%; for Model 4 (with NLP 5-grams [distinct sequences of 5 contiguous words] from the medical record), the sensitivity was 97% and the positive predictive value was 97%; and for Model 7 (a combination of Models 1 and 4), the sensitivity was 97% and the positive predictive value was 97%. Thus, NLP-based models identified 97% of surgical site infections identified by manual abstraction with high precision and 43% more surgical site infections compared with models that used administrative data only. CONCLUSIONS: Models that used NLP keywords achieved predictive abilities that were comparable with the manual abstraction process and were superior to models that used administrative data only. NLP has the potential to automate and aid accurate surgical site infection identification and, thus, play an important role in their prevention. CLINICAL RELEVANCE: This study examines NLP's potential to automate the identification of surgical site infections. This automation can potentially aid the prevention and early identification of these surgical complications, thereby reducing their adverse clinical and economic impact.