Figure 1

Detection of MDS using RT-DC and machine learning. (A) Sketch on the left shows the RT-DC setup. A syringe pump pushes suspended cells into a microfluidic chip which has a narrow constriction channel. Within the channel, cells are deformed and captured by a high-speed camera. Images are analyzed in real-time to obtain the contour (red), the convex hull (blue), the bounding box (dashed lines), and compute seven features. For each feature, the mean, median, standard deviation, and median absolute deviation are computed, resulting in 28 features describing the distributions. Sketch on the right illustrates a random forest model, which was trained to discriminate healthy and MDS, based on the 28 distribution features. (B) Barplot shows the feature importance values of a random forest model that was trained using 28 distribution features to discriminate healthy and MDS. (C) Boxplot shows the median absolute deviation (mad) of area for healthy and MDS samples. A random forest model was trained on that single feature to distinguish healthy and MDS and the resulting decision boundary is shown in the boxplot (blue corresponds to healthy and red to MDS). Boxes show the interquartile ranges (\(IQR\)), which are defined by the 25th percentile (\({Q}_{1}\)) and the 75th percentile (\({Q}_{3}\)): \(IQR={Q}_{3}-{Q}_{1}\). Yellow lines in the boxes show the medians. Whiskers represent the range of the data (lower bound: \({Q}_{1}-1.5\cdot IQR\), upper bound: \({Q}_{3}+1.5\cdot IQR\)). (D) Histograms show the distribution of area for representative measurements of healthy (blue) and MDS (red).