Energy Efficiency of Inference Algorithms for Medical Datasets: A Green AI study (Preprint)

BACKGROUND Harnessing artificial intelligence (AI) in medical domain has raised considerable interest recently. An AI model must be energy-efficient if it has to be used for inference applications in medical domain. Different from other type of data in visual AI, data in medical domain are usually composed of features with strong signals. Numerous energy optimization techniques have been developed to relieve the burden on the hardware required to deploy a complex learning model. However, the energy efficiencies of different AI models used for medical applications have not yet been studied. OBJECTIVE To explore and compare the energy efficiencies of widely-used machine learning (ML) algorithms, including logistic regression (LR), k-nearest neighbors (kNN), support vector machine (SVM), random forest (RF), extreme gradient boosting (XGB), and two different neural networks (NN) in the medical datasets. METHODS We applied the algorithms above to two distinct medical datasets, the mass spectrometry data of Staphylococcus aureus for predicting methicillin-resistance (“Mass spectrometry” dataset: 3338 cases; 268 features), and the urinalysis data for predicting Trichomonas vaginalis infection (“Urinalysis” dataset: 839,164 cases; 9 features). We compared the performance among these seven inference algorithms across accuracy, area under the receiver operating characteristic (AUROC), time consumption, and power consumption. The time and power consumptions were determined using the performance counter data from Intel Power Gadget 3.5. RESULTS Experimental results showed that the RF and XGB algorithms achieved the two highest AUROC scores with both datasets (84.7% and 83.9% with the “Mass spectrometry” dataset, respectively, and 91.1% and 91.4% with the “Urinalysis” dataset, respectively). In terms of time consumption, the XGB, 1-hidden-layer NN and LR algorithms exhibit the lowest time consumption with both datasets. RF as the referral baseline, XGB, 1-hidden-layer NN and LR achieved 45% reduction of inferencing time with the “Mass spectrometry” dataset, and 53-60% reduction with the “Urinalysis” dataset, respectively. In terms of energy efficiency, XGB, LR, SVM and RF consumed the least power. 5-hidden-layer NN as the referral baseline, XGB, LR, SVM and RF achieved 24-32% reduction of power consumption with the “Mass spectrometry” dataset, and 20-53% reduction with the “Urinalysis” dataset, respectively. Among all experiments, XGB achieved the best performance across accuracy, runtime, and energy efficiency. CONCLUSIONS In current study, XGB attained a balanced performance across accuracy, runtime, and energy efficiency in the medical datasets. The research results indicate that the XGB would be an ideal algorithm for applying ML to real-world medical scenarios.