Automated, machine-learning analysis of microclots imaged in a microfluidic device. The pipeline locates the central channel, segments the clots inside it, and turns their shapes into per-image features. A logistic regression model then uses those features to tell Long COVID (LC) and healthy control (HC) samples apart. Across 42 LC and 14 HC patients it reaches roughly 80% accuracy as a proof of principle.
Under construction. Full installation and usage instructions will follow shortly.
The central channel is isolated with an edge-based estimate followed by a flood-fill that stops where the channel meets the funnels; the outline is then fit to a clean rectangle. A neural network segments individual clots inside the channel crop, and per-clot shape statistics are read off the result.
Figure 1. (A) Raw device image. (B) Thresholded edge map. (C) Area added per flood-fill step. (D) Rectangular channel mask. (E) Channel close-up. (F) The same crop, segmented. (G) Assembly of the classifier input.
Clot features feed a logistic regression classifier that labels each sample LC or HC, with feature subsets chosen by Optuna under leave-one-out cross-validation.
Figure 2. Logistic regression pipeline and feature selection.
The network used here differs slightly from the base architecture; the actual one will be released after legal approval. Microclot segmentation labels are available on request.
This work was led by Massachusetts General Hospital (MGH).
Kirandeep K. Gill, Laurids Stockert, Bryan Alvarez-Carcamo, Jaime Greatorex, Carly Steifman, Zoe Swank, David R. Walt, Michael B. VanElzakker, Lael M. Yonker, Daniel Irimia*
*Corresponding author: Daniel Irimia.
A preprint will follow shortly.