(A-192) Title: Fostering Creativity in AI: Insights from a Datathon Tackling HIV Screening Challenges

Background: An estimated 1.2 million people in the United States are living with HIV (PWH), with up to 13% unaware of their infection and in need of testing. Identifying factors associated with HIV diagnosis could enhance screening efficiency and enable earlier treatment, reducing the risk of serious non-AIDS health outcomes and AIDS by 57%.

Objectives: To explore how a company-wide Datathon can spark innovative AI solutions that address important public health issues by fostering creativity and collaboration among interdisciplinary data scientists while prioritizing evidence-based prediction accuracy and diverse methodologies.

Methods: The Datathon utilized HealthVerity claims data linked with lab results. PWH were defined as adults with at least one HIV ICD-10-CM diagnosis code on or after October 1, 2016, with at least one year of continuous enrollment before diagnosis and no prior HIV-related records. Non-HIV controls were a 1% random sample of adults meeting similar enrollment criteria.
Participants received medical and pharmacy claims and lab results and were encouraged to augment electronic health record-based models with publicly available sociodemographic data, such as the American Community Survey. Half of the dataset was for training, with the remainder withheld for evaluation. Metrics included the area under the receiver operating characteristic (ROC) curve and weighted F1 score. A panel judged creativity and innovation based on originality, problem-solving, scientific rationale, and potential clinical impact.

Results: The dataset included 91,998 PWH and 203,083 controls, comprising 23 million medical claims, 3 million pharmacy claims, and 1 million lab results from 2016 - 2024, split into training and evaluation sets.
Twenty-eight teams with 76 participants joined the Datathon, and 19 teams submitted predictions. The top five models achieved an AUC > 0.86 and a weighted F1 score > 0.80. The high-performing models included XGBoost, transformer-based models, and model ensembles. Top predictors for delayed HIV diagnosis included history of injection drug/substance abuse, age, gender, race, geographical location, and prior sexually transmitted disease diagnoses or testing.

Conclusions: The Datathon challenged participating data scientists to apply advanced analytics to real-world data while sparking innovative approaches which included depicting medical journeys through transformer models, leveraging GenAI to identify patterns in the description of diagnosis codes, and use of ensemble models to capitalize on submodel strengths. The Datathon initiative demonstrated the potential of advanced analytics and interdisciplinary collaboration to tackle a current public health challenge.