A global taxonomy of interpretable AI: unifying the terminology for the technical and social sciences.

Since its emergence in the 1960s, Artificial Intelligence (AI) has grown to conquer many technology products and their fields of application. Machine learning, as a major part of the current AI solutions, can learn from the data and through experience to reach high performance on various tasks. This growing success of AI algorithms has led to a need for interpretability to understand opaque models such as deep neural networks. Various requirements have been raised from different domains, together with numerous tools to debug, justify outcomes, and establish the safety, fairness and reliability of the models. This variety of tasks has led to inconsistencies in the terminology with, for instance, terms such as interpretable, explainable and transparent being often used interchangeably in methodology papers. These words, however, convey different meanings and are "weighted" differently across domains, for example in the technical and social sciences. In this paper, we propose an overarching terminology of interpretability of AI systems that can be referred to by the technical developers as much as by the social sciences community to pursue clarity and efficiency in the definition of regulations for ethical and reliable AI development. We show how our taxonomy and definition of interpretable AI differ from the ones in previous research and how they apply with high versatility to several domains and use cases, proposing a-highly needed-standard for the communication among interdisciplinary areas of AI.

Partial Multiple Imputation With Variational Autoencoders: Tackling Not at Randomness in Healthcare Data.

Missing data can pose severe consequences in critical contexts, such as clinical research based on routinely collected healthcare data. This issue is usually handled with imputation strategies, but these tend to produce poor and biased results under the Missing Not At Random (MNAR) mechanism. A recent trend that has been showing promising results for MNAR is the use of generative models, particularly Variational Autoencoders. However, they have a limitation: the imputed values are the result of a single sample, which can be biased. To tackle it, an extension to the Variational Autoencoder that uses a partial multiple imputation procedure is introduced in this work. The proposed method was compared to 8 state-of-the-art imputation strategies, in an experimental setup with 34 datasets from the medical context, injected with the MNAR mechanism (10% to 80% rates). The results were evaluated through the Mean Absolute Error, with the new method being the overall best in 71% of the datasets, significantly outperforming the remaining ones, particularly for high missing rates. Finally, a case study of a classification task with heart failure data was also conducted, where this method induced improvements in 50% of the classifiers.

Male breast cancer: Looking for better prognostic subgroups.

PURPOSE: Male Breast Cancer (MBC) remains a poor understood disease. Prognostic factors are not well established and specific prognostic subgroups are warranted. PATIENTS/METHODS: Retrospectively revision of 111 cases treated in the same Cancer Center. Blinded-central pathological revision with immunohistochemical (IHQ) analysis for estrogen (ER), progesterone (PR) and androgen (AR) receptors, HER2, ki67 and p53 was done. Cox regression model was used for uni/multivariate survival analysis. Two classifications of Female Breast Cancer (FBC) subgroups (based in ER, PR, HER2, 2000 classification, and in ER, PR, HER2, ki67, 2013 classification) were used to achieve their prognostic value in MBC patients. Hierarchical clustering was performed to define subgroups based on the six-IHQ panel. RESULTS: According to FBC classifications, the majority of tumors were luminal: A (89.2%; 60.0%) and B (7.2%; 35.8%). Triple negative phenotype was infrequent (2.7%; 3.2%) and HER2 enriched, non-luminal, was rare (≤1% in both). In multivariate analysis the poor prognostic factors were: size >2 cm (HR:1.8; 95%CI:1.0-3.4 years, p = 0.049), absence of ER (HR:4.9; 95%CI:1.7-14.3 years, p = 0.004) and presence of distant metastasis (HR:5.3; 95%CI:2.2-3.1 years, p < 0.001). FBC subtypes were independent prognostic factors (p = 0.009, p = 0.046), but when analyzed only luminal groups, prognosis did not differ regardless the classification used (p > 0.20). Clustering defined different subgroups, that have prognostic value in multivariate analysis (p = 0.005), with better survival in ER/PR+, AR-, HER2-and ki67/p53 low group (median: 11.5 years; 95%CI: 6.2-16.8 years) and worst in PR-group (median:4.5 years; 95%CI: 1.6-7.8 years). CONCLUSION: FBC subtypes do not give the same prognostic information in MBC even in luminal groups. Two subgroups with distinct prognosis were identified in a common six-IHQ panel. Future studies must achieve their real prognostic value in these patients.

A new cluster-based oversampling method for improving survival prediction of hepatocellular carcinoma patients.

Liver cancer is the sixth most frequently diagnosed cancer and, particularly, Hepatocellular Carcinoma (HCC) represents more than 90% of primary liver cancers. Clinicians assess each patient's treatment on the basis of evidence-based medicine, which may not always apply to a specific patient, given the biological variability among individuals. Over the years, and for the particular case of Hepatocellular Carcinoma, some research studies have been developing strategies for assisting clinicians in decision making, using computational methods (e.g. machine learning techniques) to extract knowledge from the clinical data. However, these studies have some limitations that have not yet been addressed: some do not focus entirely on Hepatocellular Carcinoma patients, others have strict application boundaries, and none considers the heterogeneity between patients nor the presence of missing data, a common drawback in healthcare contexts. In this work, a real complex Hepatocellular Carcinoma database composed of heterogeneous clinical features is studied. We propose a new cluster-based oversampling approach robust to small and imbalanced datasets, which accounts for the heterogeneity of patients with Hepatocellular Carcinoma. The preprocessing procedures of this work are based on data imputation considering appropriate distance metrics for both heterogeneous and missing data (HEOM) and clustering studies to assess the underlying patient groups in the studied dataset (K-means). The final approach is applied in order to diminish the impact of underlying patient profiles with reduced sizes on survival prediction. It is based on K-means clustering and the SMOTE algorithm to build a representative dataset and use it as training example for different machine learning procedures (logistic regression and neural networks). The results are evaluated in terms of survival prediction and compared across baseline approaches that do not consider clustering and/or oversampling using the Friedman rank test. Our proposed methodology coupled with neural networks outperformed all others, suggesting an improvement over the classical approaches currently used in Hepatocellular Carcinoma prediction models.

Missing data imputation on the 5-year survival prediction of breast cancer patients with unknown discrete values.

Breast cancer is the most frequently diagnosed cancer in women. Using historical patient information stored in clinical datasets, data mining and machine learning approaches can be applied to predict the survival of breast cancer patients. A common drawback is the absence of information, i.e., missing data, in certain clinical trials. However, most standard prediction methods are not able to handle incomplete samples and, then, missing data imputation is a widely applied approach for solving this inconvenience. Therefore, and taking into account the characteristics of each breast cancer dataset, it is required to perform a detailed analysis to determine the most appropriate imputation and prediction methods in each clinical environment. This research work analyzes a real breast cancer dataset from Institute Portuguese of Oncology of Porto with a high percentage of unknown categorical information (most clinical data of the patients are incomplete), which is a challenge in terms of complexity. Four scenarios are evaluated: (I) 5-year survival prediction without imputation and 5-year survival prediction from cleaned dataset with (II) Mode imputation, (III) Expectation-Maximization imputation and (IV) K-Nearest Neighbors imputation. Prediction models for breast cancer survivability are constructed using four different methods: K-Nearest Neighbors, Classification Trees, Logistic Regression and Support Vector Machines. Experiments are performed in a nested ten-fold cross-validation procedure and, according to the obtained results, the best results are provided by the K-Nearest Neighbors algorithm: more than 81% of accuracy and more than 0.78 of area under the Receiver Operator Characteristic curve, which constitutes very good results in this complex scenario.

Using Kalman filters to reduce noise from RFID location system.

Nowadays, there are many technologies that support location systems involving intrusive and nonintrusive equipment and also varying in terms of precision, range, and cost. However, the developers some time neglect the noise introduced by these systems, which prevents these systems from reaching their full potential. Focused on this problem, in this research work a comparison study between three different filters was performed in order to reduce the noise introduced by a location system based on RFID UWB technology with an associated error of approximately 18 cm. To achieve this goal, a set of experiments was devised and executed using a miniature train moving at constant velocity in a scenario with two distinct shapes-linear and oval. Also, this train was equipped with a varying number of active tags. The obtained results proved that the Kalman Filter achieved better results when compared to the other two filters. Also, this filter increases the performance of the location system by 15% and 12% for the linear and oval paths respectively, when using one tag. For a multiple tags and oval shape similar results were obtained (11-13% of improvement).