Five Strategies for Bias Estimation in Artificial Intelligence-based Hybrid Deep Learning for Acute Respiratory Distress Syndrome COVID-19 Lung Infected Patients using AP(ai)Bias 2.0: A Systematic Review

Suri J. S.;Agarwal S.;Jena B.;Saxena S.;El-Baz A.;Agarwal V.;Kalra M. K.;Saba L.;Viskovic K.;Fatemi M.;Naidu S.

2022-01-01

---

Abstract

Coronavirus 2019 (COVID-19) has led to a global pandemic infecting 224 million people and has caused 4.6 million deaths. Nearly 80 Artificial Intelligence (AI) articles have been published on COVID-19 diagnosis. The first systematic review on the Deep Learning (DL)-based paradigm for COVID-19 diagnosis was recently published by Suri et al. [IEEE J Biomed Health Inform. 2021]. The above study used AtheroPoint’s “AP(ai)Bias 1.0” using 10 AI attributes in the DL framework. The proposed study uses “AP(ai)Bias 2.0” as part of the three quantitative paradigms for Risk-of-Bias quantification by using the best 40 dedicated Hybrid DL (HDL) studies and utilizing 39 AI attributes. In the first method, the radial-bias map (RBM) was computed for each AI study, followed by the computation of bias value. In the second method, the regional-bias area (RBA) was computed by the area difference between the best and the worst AI performing attributes. In the third method, ranking-bias score (RBS) was computed, where AI-based cumulative scores were computed for all the 40 studies. These studies were ranked, and the cutoff was determined, categorizing the HDL studies into three bins: low, moderate, and high. Using the Venn diagram, these three quantitative methods were benchmarked against the two qualitative non-randomized-based AI trial methods (ROBINS-I and PROBAST). Using the analytically derived moderate-high and low-moderate cutoff of 2.9 and 3.6, respectively, we observed 40%, 27.5%, 17.5%, 10%, and 20% of studies were low-biased for RBM, RBA, RBS, ROBINS-I, and PROBAST, respectively. We present an eight-point recommendation for AP(ai)Bias 2.0 minimization.