A Machine Learning Approach to Microcalorimetric Pattern Classification of Pathogens in Synovial Fluid

Figures

• 

  Figure 1. Study design. Heat flow curves were recorded by means of isothermal microcalorimetry. Heat curves were obtained as the cumulative integral of heat flow curves. Ten features were extracted from heat flow and heat curves, and used to implement machine learning (ML) classifiers to detect periprosthetic joint infection (PJI) and to distinguish between the following bacteria strains in PJI samples: Escherichia coli (EC), Enterococcus faecalis (EF), Pseudomonas aeruginosa (PA), Staphylococcus aureus (SA), and Staphylococcus epidermidis (SE). Two transfer learning convolutional neural network (CNN) models were also implemented to distinguish between bacteria strains, using scalograms and images of heat flow curves as input data. Probability distributions of the two transfer learning CNN models were averaged to obtain a combined transfer learning CNN model for PJI identification.

• 

  Figure 2. Swarm scatter chart of growth parameters. Inflection time (t i ), upper horizontal asymptote (A), lag time (t 0), linear period (T lin), threshold time (T thr), maximum heat flow (V i ), maximum metabolic velocity (MMV), maximum metabolic velocity time (MMV t ), maximum decay velocity (MDV), and maximum decay velocity time (MDV t ) were used to characterize heat flow and heat curves obtained from aseptic samples (NEG) and the following PJI samples: Escherichia coli (EC), Enterococcus faecalis (EF), Pseudomonas aeruginosa (PA), Staphylococcus aureus (SA), and Staphylococcus epidermidis (SE).

• 

  Figure 3. Performance metrics of the multiclass XGBoost classifier for PJI identification. Escherichia coli (EC), Enterococcus faecalis (EF), Pseudomonas aeruginosa (PA), Staphylococcus aureus (SA), Staphylococcus epidermidis (SE), precision (Prec), recall (Rec), F1‐score (F1), area under the curve (AUC), and accuracy (Acc). Precision, recall, F1‐score, and accuracy metrics are expressed as percentages.

• 

  Figure 4. Performance metrics of the binary XGBoost classifier for PJI detection. Precision (Prec), recall (Rec), F1‐score (F1), area under the curve (AUC), and accuracy (Acc). Precision, recall, F1‐score, and accuracy metrics are expressed as percentages.

• 

  Figure 5. Performance metrics of the combined transfer learning CNN classifier for PJI identification. Escherichia coli (EC), Enterococcus faecalis (EF), Pseudomonas aeruginosa (PA), Staphylococcus aureus (SA), Staphylococcus epidermidis (SE), precision (Prec), recall (Rec), F1‐score (F1), area under the curve (AUC), and accuracy (Acc). Precision, recall, F1‐score, and accuracy metrics are expressed as percentages.

• 

  Figure 6. Impact of bacterial growth parameters in the multiclass XGBoost classifier for PJI identification. Upper horizontal asymptote (A), maximum metabolic velocity (MMV), maximum decay velocity (MDV), maximum decay velocity time (MDV t ), maximum heat flow (V i ), inflection time (t i ), threshold time (T thr), maximum metabolic velocity time (MMV t ), linear period (T lin), and lag time (t 0) were calculated from heat flow and heat curves of the following PJI samples: Escherichia coli (EC), Enterococcus faecalis (EF), Pseudomonas aeruginosa (PA), Staphylococcus aureus (SA), and Staphylococcus epidermidis (SE).

• 

  Figure 7. Local importance of bacterial growth parameters from two Pseudomonas aeruginosa (PA) samples in the multiclass XGBoost classifier. Percentage values in the top panel represent the probability of belonging to the PA class, as returned by the XGBoost classifier. Maximum heat flow (V i ), inflection time (t i ), threshold time (T thr), upper horizontal asymptote (A), lag time (t 0), linear period (T lin), maximum metabolic velocity (MMV), maximum metabolic velocity time (MMV t ), maximum decay velocity (MDV), and maximum decay velocity time (MDV t ).