Optimizing Obesity Prediction Models with Green AI: A Multi-Criteria Approach
Enobong Imo Udo
Department of Computer Engineering, University of Uyo, Nwaniba Road, Uyo, 270001, Akwa Ibom State, Nigeria.
Bliss Utibe-Abasi Stephen *
Department of Computer Engineering, University of Uyo, Nwaniba Road, Uyo, 270001, Akwa Ibom State, Nigeria.
Philip Asuquo
Department of Computer Engineering, University of Uyo, Nwaniba Road, Uyo, 270001, Akwa Ibom State, Nigeria.
Ihemereze Chijioke Nnanna
Department of Computer Engineering, Federal Polytechnic, Nekede, 460109, Imo State, Nigeria.
Amarachi Vivian Nnajiofor
Department of Computer Engineering, University of Uyo, Nwaniba Road, Uyo, 270001, Akwa Ibom State, Nigeria.
Michael Okon Bassey
Department of Mechatronics Engineering, Akwa Ibom State Polytechnic, Ikot Ekpene, 530101, Akwa Ibom State, Nigeria.
Iniubongabasi Paul Etim
Bridge Clinic, Umaru Dikko St, Jabi, Abuja 900108, Abuja, Nigeria.
*Author to whom correspondence should be addressed.
Abstract
Aims/ Objectives: With the consistent global increase in obesity population, researchers have proposed both machine learning and neural network techniques in predicting obesity beforehand from lifestyle and demographic factors. However, extant literatures in the pipeline subscribes to existing Red AI approach, that is, enhancing model’s accuracy at whatever cost without recourse to the effect of such model on the environment in terms of carbon footprint and efficiency.
Study Design: This paper set forward a green AI model selection strategy for obesity risk prediction which depends on multi-criteria determination method, factoring in computation time, in choosing the optimal model for final deployment.
Methodology: In this research, we applied two deep learning and three machine learning models. We used the algorithms of Convolution Neural Network (CNN), Artificial Neural Network (ANN), Random Forest (RF), Logistic Regression (LR) and Decision Trees (DT) learning algorithms. Our proposed method identifies the optimal model for final deployment based on equal compromise between model’s performance and computation time. The estimated Carbon footprint and Energy consumption of training the models used in this study have been computed using green Algorithms.
Results: From our comparative analysis, our green AI selection strategy favored Random forest model, which scored 97.16% in accuracy, took 0.03420 seconds of computation time, 2.36mg of CO2e carbon footprint and 2.62×10−3 WH energy consumption during model training and validation.
Conclusion: This paper’s contributions are significant to the support of the ongoing call for Green AI, especially within the healthcare sector. Moreover, findings imply that we do not solve the health challenge of obesity while creating others with increased carbon footprint.
Keywords: Obesity, machine learning, green AI, deep learning, carbon footprint, energy consumption