ABSTRACT
This paper studies how counterfactual explanations can be used to assess the fairness of a model. Using machine learning for high-stakes decisions is a threat to fairness as these models can amplify bias present in the dataset, and there is no consensus on a universal metric to detect this. The appropriate metric and method to tackle the bias in a dataset will be case-dependent, and it requires insight into the nature of the bias first. We aim to provide this insight by integrating explainable AI (XAI) research with the fairness domain. More specifically, apart from being able to use (Predictive) Counterfactual Explanations to detect explicit bias when the model is directly using the sensitive attribute, we show that it can also be used to detect implicit bias when the model does not use the sensitive attribute directly but does use other correlated attributes leading to a substantial disadvantage for a protected group. We call this metric PreCoF, or Predictive Counterfactual Fairness. Our experimental results show that our metric succeeds in detecting occurrences of implicit bias in the model by assessing which attributes are more present in the explanations of the protected group compared to the unprotected group. These results could help policymakers decide on whether this discrimination is justified or not.
ABSTRACT
BACKGROUND: Personalized medicine tailors care based on the patient's or pathogen's genotypic and phenotypic characteristics. An automated Clinical Decision Support System (CDSS) could help translate the genotypic and phenotypic characteristics into optimal treatment and thus facilitate implementation of individualized treatment by less experienced physicians. METHODS: We developed a hybrid knowledge- and data-driven treatment recommender CDSS. Stakeholders and experts first define the knowledge base by identifying and quantifying drug and regimen features for the prototype model input. In an iterative manner, feedback from experts is harvested to generate model training datasets, machine learning methods are applied to identify complex relations and patterns in the data, and model performance is assessed by estimating the precision at one, mean reciprocal rank and mean average precision. Once the model performance no longer iteratively increases, a validation dataset is used to assess model overfitting. RESULTS: We applied the novel methodology to develop a treatment recommender CDSS for individualized treatment of drug resistant tuberculosis as a proof of concept. Using input from stakeholders and three rounds of expert feedback on a dataset of 355 patients with 129 unique drug resistance profiles, the model had a 95% precision at 1 indicating that the highest ranked treatment regimen was considered appropriate by the experts in 95% of cases. Use of a validation data set however suggested substantial model overfitting, with a reduction in precision at 1 to 78%. CONCLUSION: Our novel and flexible hybrid knowledge- and data-driven treatment recommender CDSS is a first step towards the automation of individualized treatment for personalized medicine. Further research should assess its value in fields other than drug resistant tuberculosis, develop solid statistical approaches to assess model performance, and evaluate their accuracy in real-life clinical settings.
