Announcing The NIH Long COVID Computational Challenge (L3C) Winners

In August 2022, the Rapid Acceleration of Diagnostics Radical (RADx®-rad) program at the National Institutes of Health (NIH) launched the Long COVID Computational Challenge (L3C) to support creative data-driven solutions that meaningfully advance the current understanding of the risks of developing post-acute sequelae of SARS-Co-V2 (PASC)/Long COVID. The L3C complements the NIH’s other Long COVID research initiatives, like Researching COVID to Enhance Recovery (RECOVER).

The primary objective of the L3C was to focus on the prognostic problem by developing artificial intelligence/machine learning (AI/ML) models and algorithms that serve as open-source tools for using structured medical records to identify which patients infected with SARS-CoV-2 have a high likelihood of developing PASC/Long COVID. Challenge participants were expected to develop, train, and test their models to aid in predicting the susceptibility to and likelihood of developing PASC/Long COVID in patients with SARS-CoV-2 infection. Challenge participants utilized de-identified electronic health record (EHR) data available through NCATS’s National COVID Cohort Collaborative (N3C) Data Enclave, a central, harmonized data repository that represents EHR data from over 74 health centers across the United States.

At the close of the submission period, 74 teams had registered and onboarded to N3C, and 35 teams had completed submissions. The participating teams included scientists from universities, medical centers, industry and public/private partnerships. Complete submissions underwent a quantitative and qualitative evaluation and judges selected three winners and three honorable mentions. The original announcement can be viewed on Challenge.gov.

1st Place, $200,000 - Convalesco

Real-time prediction of PASC risks in COVID patients.

Convalesco built a real-time monitoring system based on LightGBM and XGBoost that updates a patient’s risk for developing PASC/Long COVID as new clinical events occur. Their submission visualized this cumulative risk in a Clinical Decision Support dashboard.

2nd Place, $150,000 - Geisinger AI Lab (GAIL)

A Long COVID Prediction Model

GAIL built a portable, efficient, and accurate model using fewer features than the competition, resulting in a translational PASC/Long COVID prediction clinical decision support tool that included EHR-ready summary visualizations for clinician interpretation.

3rd Place, $100,000 - University of California, Berkeley (UC Berkeley)

Predicting PASC/Long COVID in a Matched Sample of Patients with COVID-19

This team out of UC Berkeley built a clinical prediction model that combined many smaller prediction models (this combined model is known as an ensemble or a Super Learner). The model used various aspects of a patient’s health such as their cardiovascular health, respiratory health, history of hospital use, and age to predict the patient’s risk for developing PASC/Long COVID.

Honorable Mention - University of Wisconsin-Madison, Department of Biostatistics & Medical Informatics (BMI)

A LightGBM model to predict Long COVID patients

UW-Madison BMI built a PASC/Long COVID prediction model by looking at high-level clinical concepts in a patient’s clinical history to evaluate their risk of developing PASC/Long COVID. This model was fourth place overall in judging.

Honorable Mention - University of Pennsylvania (Penn)

A robust and adaptive model for Long COVID detection

Penn took a unique approach to this challenge and developed a PASC/Long COVID prediction model that looked at both static clinically relevant data points as well as dynamically selected data points. This grounded their model in clinical relevancy but allowed it to adapt to future changes in new data.

Honorable Mention - Ruvos

Comparison of machine learning and ensemble models to predict patients developing PASC/Long COVID

Ruvos developed a prediction model that used broad categories of disease to predict a patient’s risk of developing PASC/Long COVID. Their model was highly generalizable to new EHR data.