(ASPOT09) Development of a Machine Learning Prediction Model for COVID-19 Patient Outcome Using a Sequential Data-Based Patient Similarity Metric

Background: While various prognosis prediction models for COVID-19 patients have been proposed, few have utilized sequential data that consider the progression of the disease, due to their irregularity and heterogeneity.

Objectives: We aimed to develop a COVID-19 patient outcome prediction model through sequential data that reflects disease progression, based on patient similarity.

Methods: The Common Data Model from a tertiary hospital in Korea, spanning February 2020 to June 2021, was used. Patient similarity was calculated using a weighted sum from three different types of feature similarities. First, sequential similarity, derived from timestamped event sequences (20 types of symptoms; lung ultrasounds, CT scans, and X-ray tests), was calculated using the edit distance method. Second, trend similarity, based on time-series data (27 types of laboratory test results), was determined using the dynamic time warping method. Lastly, cross-sectional information, including age, sex, 27 underlying diseases, and 49 drug prescriptions, was integrated using the Jaccard distance. Baseline data refers to data collected before the diagnosis of COVID-19. Timestamped event sequences or time-series data utilize information collected within 30 days after the COVID-19 diagnosis. We developed a k-nearest neighbor (KNN) model using the patient similarity matrix to predict 90-day mortality. To compare the performance of the KNN model, we examined results from logistic regression, random forest, long short-term memory network, and recurrent neural network models.

Results: A total of 5,969 adult patients (mean age: 61.5 years; men: 54.2%) with COVID-19 who were admitted to or visited the emergency room were analyzed. Data from 762,663 lab tests (median: 69; interquartile range [IQR]: 33-144 per patient) and 28,231 symptoms, lung ultrasounds, CT scans, and X-ray tests (median: 3, IQR: 2-6 per patient) were extracted to construct a 5,969 x 5,969 similarity matrix. Our KNN model (k=50) demonstrated high accuracy with an area under the receiver operating characteristic curve of 0.79 in predicting 90-day mortality. The KNN model outperformed other machine learning techniques in predictive accuracy.

Conclusions: Our study demonstrates that patient outcomes can be accurately predicted using a patient similarity metric that considers disease progression and clinical course. This highlights the potential of such approaches for clinical management, especially during pandemic outbreaks.