Machinometrics, a new framework towards explainable machine learning in healthcare

Machine learning (ML) offers significant potential in healthcare, but its clinical implementation is hampered by the ‘black box nature’ of these algorithms. This lack of transparency necessitates explainable models to ensure that the machine’s decisions are interpretable and can be trusted by healthcare professionals. To address this issue, we introduce a novel Physically Grounded Explanation Framework using ‘Machinometrics’ that aims to make the machine’s decisions more understandable and relatable to healthcare professionals. We used the Machinometrics framework to enhance the interpretability of ML models in tooth identification tasks, a foundational application in clinical dentistry. The data set contained 1236 valid 3D CT scans of human teeth, categorised into eight classes. Data pre-processing converted DICOM files into NRRD format, isolating individual tooth volumes. The teeth data were split into five folds for cross-validation, and a 3D ResNet model was employed to identify the tooth. The model achieved an accuracy of 77.72%. Post-hoc analysis using Gradient-weighted Class Activation Mapping (Grad-CAM) was used to generate heatmap data, which were then correlated with the physical attributes such as density and distance to generate Machinometrics plots. Principal Component Analysis (PCA) was applied to further substantiate the model’s decisions and reliability. The case study on tooth classification demonstrated the potential of the Machinometrics framework in clinical applications. Density maps highlighted different tooth structures, such as enamel, dentine and pulp, while distance maps depicted the spatial relationship between tooth surfaces and surrounding space. PCA further substantiated the model’s reliability, revealing significant patterns and correlations. Machinometrics offers a promising avenue to recontextualise black-box models back to the physical world by serving as a common language among computer scientists and end users. By integrating physical attributes such as density and distance, we provided tangible, contextualised explanations of the model’s decisions through 3D visualisations. This approach allows clinicians to intuitively assess and quantitatively evaluate the algorithm’s outputs. This would, in turn, foster practical applications of advanced computational models in healthcare.