Ensemble learning by diversifying explanations: predicting the evolution of pain relief

Abstract

Modeling from data usually has two distinct facets: building sound explanatory models or creating powerful predictive models for a system or phenomenon. While there is an instinctive sense that predicting and explaining are distinct tasks, it is often assumed that models with high explanatory power are inherently of high predictive power. In spite of this relationship, most recent data-modeling work fits into the allin-one machine learning methodology, with the basic assumption that all important explanatory factors can be combined into a single predictive model. Although highly adopted and established, the all-in-one methodology neglects that many phenomena are actually defined by several subpopulations or local structures and therefore there are many possible predictive models that provide contrasting interpretations or competing explanations for the same phenomenon. In this work, we present ED-Ensemble (Explanation-Diversifying Ensemble), an alternative to the all-in-one methodology. Our main intuition is that models that have their decisions explained by the same factors will probably perform better predictions within the same local structures. We design and conduct an experimental evaluation as a case study to evaluate the performance of our methodology to model the evolution of pain relief in patients suffering from chronic pain under usual guideline-based treatment. Six hundred thirty-one participants self-completed the McGill Pain Questionnaire and the Visual Analogue Scale. Chronic pain can be primary or secondary to diseases. Its symptomatology can be classified as nociceptive, nociplastic or neuropathic, and is generally associated with many different causal structures, challenging the typical all-in-one methodology. We show that we can effectively combine models with competing explanations, promoting diversity in ensemble, leading to significant gains in accuracy by enforcing a stable scenario in which models that are similar in terms of their predictions are also similar in terms of explanatory factors. Further, we present EXP-MF (model-EXPlanations as MetaFeatures). We follow the explanation-diversity feature selection proposed and extend it to use model-explanations as meta-features in longitudinal data, as the standard protocol for a patient typically comprises many subsequent appointments. This approach prevents us from neglecting a considerable amount of information. The ED-Ensemble obtained from our experiments consistently outperformed the all-in-one approaches, notwithstanding employing state-of-art ensemble algorithms XGBoost and Random Forest. Our proposed approach considering the first consultation only achieved an AUC of 0.78 using XGBoost as learning algorithm, relative performance gain up to 20.37% compared to the XGBoost all-in-one, and AUC of 0.75 when using Random Forest as learning algorithm, relative performance gain up to 15.03% compared to the Random Forest all-in-one approach. Also, the number of features is remarkably reduced, using as low as 15% of features. When considering sequential consultations, the experiments consistently showed that the more consultations granted, the higher the performance achieved. Our approach EXP-MF with an ED-Ensemble could achieve an AUC of 0.945 (increase of 23.37%) using five consultations. A similar uptrend in AUC was also observed for the XGBoost and Random Forest algorithms, achieving 0.843 (increase of 50%) and 0.810 (increase of 62.98%) respectively. Finally, our novel ensemble based on diversified explanations presented as a superior alternative to the all-in-one approach in multiple-structure phenomena problems with cross-sectional and longitudinal data.