Energy Efficiency of Inference Algorithms for Clinical Laboratory Data Sets: Green Artificial Intelligence Study

Area of Science:

• Medical Artificial Intelligence (AI)
• Machine Learning in Healthcare
• Computational Biology

Background:

• AI is increasingly used in medicine, requiring energy-efficient models for inference.
• Medical lab data often has strong signals, but AI model energy efficiency is understudied.
• Optimization techniques exist, but comparative energy efficiency analysis of AI models for medical tasks is lacking.

Purpose of the Study:

• To compare the energy efficiency of common machine learning algorithms for clinical laboratory data.
• To evaluate logistic regression (LR), k-nearest neighbor, support vector machine, random forest (RF), extreme gradient boosting (XGB), and neural network (NN) variants.
• To assess performance based on accuracy, AUROC, inference time, and power consumption.

Main Methods:

• Applied nine inference algorithms to two distinct clinical datasets: mass spectrometry (Staphylococcus aureus) and urinalysis (Trichomonas vaginalis).
• Measured accuracy, AUROC, time, and power consumption using Intel Power Gadget 3.5.
• Compared algorithm performance on datasets with 3338 and 839,164 cases, respectively.

Main Results:

• XGB and RF achieved the highest AUROC values for both datasets.
• XGB and LR algorithms showed the shortest inference times.
• XGB demonstrated the lowest power consumption on the mass spectrometry dataset, while LR was most efficient for urinalysis.
• XGB consistently offered the best balance of accuracy, speed, and energy efficiency.

Conclusions:

• The extreme gradient boosting (XGB) algorithm provides a balanced and efficient solution for medical AI inference.
• XGB's performance in accuracy, run time, and energy efficiency makes it suitable for real-world medical applications with energy constraints.