Exploring Text Specific and Blackbox Fairness Algorithms in Multimodal Clinical NLP

Abstract

Clinical machine learning is increasingly multimodal, collected in both structured tabular formats and unstructured forms such as free text. We propose a novel task of exploring fairness on a multimodal clinical dataset, adopting equalized odds for the downstream medical prediction tasks. To this end, we investigate a modality-agnostic fairness algorithm-equalized odds post processing-and compare it to a text-specific fairness algorithm: debiased clinical word embeddings. Despite the fact that debiased word embeddings do not explicitly address equalized odds of protected groups, we show that a text-specific approach to fairness may simultaneously achieve a good balance of performance and classical notions of fairness. We hope that our paper inspires future contributions at the critical intersection of clinical NLP and fairness. The full source code is available here: https://github.com/ johntiger1/multimodal_fairness

The Road to Explainability is Paved with Bias: Measuring the Fairness of Explanations

“…There are three main strategies for encouraging group fairness [20]: pre-processing data to find less-biased representations [81]; enforcing fairness while training a model, typically through regularization [72,118]; and altering a model's predictions to satisfy fairness constraints after it is trained [2,24,48,83]. In this paper, we utilize the inprocessing method proposed by Zhang et al [118] for training fair blackbox models.…”

Broadening AI Ethics Narratives: An Indic Art View

“…This not only entails providing multimodal explanations underlying an AI decision and training systems through demonstrations, but also leveraging multimodal data sources (text, image, video, audio) encompassing diverse ethical perspectives in defining, designing, and developing ethical AI systems. Existing works only address parts of the proposed strategy such as through multimodal explanations [41,63], exploration of fairness related issues in multimodal settings [7,69], inference of norms from stories [15], or by using principles of imitation learning towards value alignment [61] ; these works neither leverage multimodal data sources nor incorporate diverse ethical perspectives in their design. For example, oral traditions of teaching in Indian art education are heavily informed by "Dhvani" or sound signals.…”

MIMIC-IF: Interpretability and Fairness Evaluation of Deep Learning Models on MIMIC-IV Dataset

“…With clinical notes 49,50 or temporal measurements 4,51,52 or both 53 from MIMIC-III considered, fairness evaluation and bias mitigation have been studied recently for tasks such as mortality prediction 4,[49][50][51][52][53] , phenotyping 50,53 , readmission 51 , length of stay 52 , etc. To evaluate data and prediction fairness for the aforementioned healthcare tasks, attributes like ethnicity 4,49,50,52,53 , gender 50,52,53 , insurance 50,53 , age 49 and language 50 , are considered most often to split patients into different protected groups.…”

“…When making medical decisions based on text data like clinical notes, word embeddings, used as machine learning inputs, have been demonstrated to propagate unwanted relationships with regard to different genders, language speakers, ethnicities, and insurance groups 50,53 . With respect to gender and insurance type, differences in accuracy and therefore machine bias has been observed for mortality prediction 51 .…”

“…Martinez et al proposed an in-processing approach where the fairness problem is characterized as a multi-objective optimization task, where the risk for each protected group is a separate objective 49 . After well-trained machine learning models make predictions, equalized odds post-processing 53 and updating predictions according to the weighted sum of utility and fairness 52 were introduced respectively as effective post-processing approaches.…”