Pedro Henriques Abreu

Abstract
Missing data is an issue that can negatively impact any task performed with the available data and it is often found in real -world domains such as healthcare. One of the most common strategies to address this issue is to perform imputation, where the missing values are replaced by estimates. Several approaches based on statistics and machine learning techniques have been proposed for this purpose, including deep learning architectures such as generative adversarial networks and autoencoders. In this work, we propose a novel siamese neural network suitable for missing data imputation, which we call Siamese Autoencoder-based Approach for Imputation (SAEI). Besides having a deep autoencoder architecture, SAEI also has a custom loss function and triplet mining strategy that are tailored for the missing data issue. The proposed SAEI approach is compared to seven state-of-the-art imputation methods in an experimental setup that comprises 14 heterogeneous datasets of the healthcare domain injected with Missing Not At Random values at a rate between 10% and 60%. The results show that SAEI significantly outperforms all the remaining imputation methods for all experimented settings, achieving an average improvement of 35%. This work is an extension of the article Siamese Autoencoder-Based Approach for Missing Data Imputation [1] presented at the International Conference on Computational Science 2023. It includes new experiments focused on runtime, generalization capabilities, and the impact of the imputation in classification tasks, where the results show that SAEI is the imputation method that induces the best classification results, improving the F1 scores for 50% of the used datasets.

Abstract
Experimental assessment of different missing data imputation methods often compute error rates between the original values and the estimated ones. This experimental setup relies on complete datasets that are injected with missing values. The injection process is straightforward for the Missing Completely At Random and Missing At Random mechanisms; however, the Missing Not At Random mechanism poses a major challenge, since the available artificial generation strategies are limited. Furthermore, the studies focused on this latter mechanism tend to disregard a comprehensive baseline of state-of-the-art imputation methods. In this work, both challenges are addressed: four new Missing Not At Random generation strategies are introduced and a benchmark study is conducted to compare six imputation methods in an experimental setup that covers 10 datasets and five missingness levels (10% to 80%). The overall findings are that, for most missing rates and datasets, the best imputation method to deal with Missing Not At Random values is the Multiple Imputation by Chained Equations, whereas for higher missingness rates autoencoders show promising results.

Abstract
Medical imaging classification improves patient prognoses by providing information on disease assessment, staging, and treatment response. The high demand for medical imaging acquisition requires the development of effective classification methodologies, occupying deep learning technologies, the pool position for this task. However, the major drawback of such techniques relies on their black-box nature which has delayed their use in real-world scenarios. Interpretability methodologies have emerged as a solution for this problem due to their capacity to translate black-box models into clinical understandable information. The most promising interpretability methodologies are concept-based techniques that can understand the predictions of a deep neural network through user-specified concepts. Concept activation regions and concept activation vectors are concept-based implementations that provide global explanations for the prediction of neural networks. The explanations provided allow the identification of the relationships that the network learned and can be used to identify possible errors during training. In this work, concept activation vectors and concept activation regions are used to identify flaws in neural network training and how this weakness can be mitigated in a human-in-the-loop process automatically improving the performance and trustworthiness of the classifier. To reach such a goal, three phases have been defined: training baseline classifiers, applying the concept-based interpretability, and implementing a human-in-the-loop approach to improve classifier performance. Four medical imaging datasets of different modalities are included in this study to prove the generality of the proposed method. The results identified concepts in each dataset that presented flaws in the classifier training and consequently, the human-in-the-loop approach validated by a team of 2 clinicians team achieved a statistically significant improvement.

Abstract

Abstract