Axes of Prognosis: Identifying Subtypes of COVID-19 Outcomes

Abstract

COVID-19 is a disease with vast impact, yet much remains unclear about patient outcomes. Most approaches to risk prediction of COVID-19 focus on binary or tertiary severity outcomes, despite the heterogeneity of the disease. In this work, we identify heterogeneous subtypes of COVID-19 outcomes by considering 'axes' of prognosis. We propose two innovative clustering approaches - 'Layered Axes' and 'Prognosis Space' - to apply on patients' outcome data. We then show how these clusters can help predict a patient's deterioration pathway on their hospital admission, using random forest classification. We illustrate this methodology on a cohort from Wuhan in early 2020. We discover interesting subgroups of poor prognosis, particularly within respiratory patients, and predict respiratory subgroup membership with high accuracy. This work could assist clinicians in identifying appropriate treatments at patients' hospital admission. Moreover, our method could be used to explore subtypes of 'long COVID' and other diseases with heterogeneous outcomes.

Figure 1:

The system architecture to discover and predict COVID-19 subtypes.

Figure 2:

Two approaches to clustering using axes of prognosis.

Figure 3:

Heatmaps showing prevalence of binary features within each cluster found by each approach. OXY refers to use of oxygen therapy and LOS refers to the length of stay. Note that the features shown in the third row, Demographic information, were not used for clustering - they are shown only to aid interpretation of the clusters. For the Layered Axes K-Modes clusters, labels are in the form number-letter where the number indicates the clusters found from the first axis, and the letter the subclusters found within that cluster using features from the second axis. Note also that for the Prognosis Space approach, DBSCAN has been used and -1 indicated points described as ‘noise’.

Figure 4:

Example confusion matrices heatmaps for Layered Axes K-Modes classification subclusters. Left: Nutritional; Middle: Respiratory; Right: Circulatory. x-axis: predicted label; y-axis: true label.

Figure 5:

Example confusion matrices for classification of severe patients. Left: Layered Axes K-Modes; Right: Prognosis Space DBSCAN. x-axis: predicted label; y-axis: true label.