Addressing Consequential Public Health Problems Through Informatics and Data Science

In 2020 and 2021, Journal of the American Medical Informatics Association (JAMIA) published many papers related to the COVID-19 pandemic. Despite the COVID-19 pandemic’s prominence as a public health issue, other consequential public problems continue to plague our society and, in some instances, have been exacerbated by the pandemic. In this editorial, I highlight papers related to suicide, opioid use disorder, and child abuse.

Given that accurate identification of self-harm presentations to Emergency Departments (EDs) can lead to more timely mental health support, aid in understanding the burden of suicidal intent in a population, and support evaluation of public health initiatives related to suicide prevention, Rozova et al¹ developed an automated system for the detection of self-harm presentations from brief ED nursing triage notes. They applied natural language processing to 477 627 free-text triage notes from ED presentations in a single site; 1.4% were labeled as related to self-harm by 2 annotators. They evaluated the performance of multiple machine learning models finding that the calibrated Gradient Boosting model had the best performance for classifying the presence of self-harm in ED triage notes. Although a major limitation was that the ED triage notes were from a single site, the findings show promise for identifying patients who would benefit from mental health follow-up as well as for supporting population surveillance of self-harm and evaluating the impact of suicide prevention efforts.

Also on the topic of suicide, Xu et al² proposed a general targeted fusion learning framework for building tailored risk prediction models. They fused a limited statewide Hospital Inpatient Discharge Dataset (HIDD) with the more comprehensive medical All-Payer Claims Database (APCD) from Connecticut to predict suicide-related hospitalizations for pediatric patients. They built a suicide risk prediction model for the more comprehensive source data (APCD) and calculated patient risk scores. They created patient similarity scores between patients in the source and target (HIDD) datasets based on demographic characteristics and diagnosis codes and then generated risk score for each patient in the target dataset. They found that the positive predictive values for the combined fusion models had superior positive predictive values compared to the conventional model. Their findings suggest that the general target fusion learning framework including patient similarity scores can be used to improve the performance of predictive models in a more limited dataset by taking advantage of a more comprehensive data set.

Varshney et al³ derived the RACE (Review, Assess, Classify, and Evaluate) framework for analyzing mobile health (mHealth) apps by integrating existing methodologies. In step 1, Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) is used to systematically review the literature and app platforms to identify the apps for the condition of interest. In step 2, inter-rater reliability analysis and the Nickerson-Varshney-Muntermann taxonomy are used to assess and classify the apps’ specific dimensions and characteristics. Step 3 involves evaluating the apps according to the step 2 taxonomy and evidence-based criteria to identify patterns and functionality gaps that can inform future research and mHealth design for the topic of interest. They applied RACE to examine mHealth apps focused on the critical public health issue of opioid use disorder. Among 153 opioid apps, the target audience was predominantly patients and healthcare professionals, and the most common functions were treatment and recovery (37%), education (24%), prescription (16%), and reminder-monitoring-support (13%). Security and privacy protection was only evident in 84% of the apps. The findings support the use of the RACE framework as a strategy to inform understanding of the dimensions and characteristics of mHealth apps for a specific health condition.

Landau et al^4,5 addressed the vital public health topic of child abuse and neglect through a qualitative study and a related Perspective. They interviewed 20 pediatric clinicians working in a single pediatric emergency department (ED) to gain insights about generating an electronic health record (EHR)-based phenotype to identify children at risk for abuse and neglect.⁴ Their analysis identified 3 central themes: (a) challenges in diagnosing child abuse and neglect (eg, inadequate previous visit history); (b) health discipline differences in documentation styles (eg, in different document types) in the EHR; and (c) identification of potential racial bias through documentation (eg, judgmental statements about level of clinician belief in account of injury). Study findings will inform the development of an EHR phenotype and clinical decision support that minimizes racial biases in the identification of child abuse and neglect. In the Perspective, the authors contend that using EHR data and machine learning to develop and evaluate risk models for the detection of child abuse and neglect raises critical issues.⁵ Issues related to the practicality of machine learning-based risk models due to limitations in the available data include (a) lack of centralized evidence base for child abuse and neglect; (b) racial biases in EHR data; and (c) clinical documentation system design issues. From the ethical issues standpoint, there are challenges in the evaluation of risk prediction performance (eg, errors have severe consequences for those identified through risk models), testing predictive models in the real world (eg, a randomized controlled trial may be considered unethical), and presentation of prediction results to clinicians (eg, some models are not explainable). To address these ethical challenges, the authors recommend external validation of models by individuals not involved with model development, use of research designs other than randomized controlled trials to test the models in practice, and the initial development of models that are transparent and can be interpreted by clinicians.

While JAMIA will continue to publish high-quality and innovative content related to the COVID-19 pandemic, we remain committed to publishing papers that apply informatics and data science to other consequential public health issues such as those highlighted in this editorial suicide, opioid use disorder, and child abuse and neglect.

CONFLICT OF INTEREST STATEMENT

None.