Laura Arnal

The rapid adoption of complex Artificial Intelligence (AI) and Machine Learning (ML) models has led to their characterization as black boxes due to the difficulty of explaining their internal decision-making processes. This lack of transparency hinders users' ability to understand, validate and trust model behavior, particularly in high-risk applications. Although explainable AI (XAI) has made significant progress, there remains a need for versatile and effective techniques to address increasingly complex models. This work introduces Multivariate Conditional Expectation (MUCE), a model-agnostic method for local explainability designed to capture prediction changes from feature interactions. MUCE extends Individual Conditional Expectation (ICE) by exploring a multivariate grid of values in the neighborhood of a given observation at inference time, providing graphical explanations that illustrate the local evolution of model predictions. In addition, two quantitative indices, stability and uncertainty, summarize local behavior and assess model reliability. Uncertainty is further decomposed into uncertainty+ and uncertainty- to capture asymmetric effects that global measures may overlook. The proposed method is validated using XGBoost models trained on three datasets: two synthetic (2D and 3D) to evaluate behavior near decision boundaries, and one transformed real-world dataset to test adaptability to heterogeneous feature types. Results show that MUCE effectively captures complex local model behavior, while the stability and uncertainty indices provide meaningful insight into prediction confidence. MUCE, together with the ICE modification and the proposed indices, offers a practical contribution to local explainability, enabling both graphical and quantitative insights that enhance the interpretability of predictive models and support more trustworthy and transparent decision-making.

Type 2 diabetes (T2D) is becoming one of the leading health problems in Western societies, diminishing quality of life and consuming a significant share of healthcare resources. This study presents machine learning models for T2D diagnosis and prognosis, developed using heterogeneous data from a Spanish population dataset (Di@bet.es study). The models were trained exclusively on individuals classified as controls and undiagnosed diabetics, ensuring that the results are not influenced by treatment effects or behavioral changes due to disease awareness. Two data domains are considered: environmental (patient lifestyle questionnaires and measurements) and clinical (biochemical and anthropometric measurements). The preprocessing pipeline consists of four key steps: geospatial data extraction, feature engineering, missing data imputation, and quasi-constancy filtering. Two working scenarios (Environmental and Healthcare) are defined based on the features used, and applied to two targets (diagnosis and prognosis), resulting in four distinct models. The feature subsets that best predict the target have been identified based on permutation importance and sequential backward selection, reducing the number of features and, consequently, the cost of predictions. In the Environmental scenario, models achieved an AUROC of 0.86 for diagnosis and 0.82 for prognosis. The Healthcare scenario performed better, with an AUROC of 0.96 for diagnosis and 0.88 for prognosis. A partial dependence analysis of the most relevant features is also presented. An online demo page showcasing the Environmental and Healthcare T2D prognosis models is available upon request. Graphic

Missing values are a major challenge in most data science projects working on real data. To avoid losing valuable information, imputation methods are used to fill in missing values with estimates, allowing the preservation of samples or variables that would otherwise be discarded. However, if the process is not well controlled, imputation can generate spurious values that introduce uncertainty and bias into the learning process. The abundance of univariate and multivariate imputation techniques, along with the complex trade-off between data reliability and preservation, makes it difficult to determine the best course of action to tackle missing values. In this work, we present ITI-IQA (Imputation Quality Assessment), a set of utilities designed to assess the reliability of various imputation methods, select the best imputer for any feature or group of features, and filter out features that do not meet quality criteria. Statistical tests are conducted to evaluate the suitability of every tested imputer, ensuring that no new biases are introduced during the imputation phase. The result is a trainable pipeline of filters and imputation methods that streamlines the process of dealing with missing data, supporting different data types: continuous, discrete, binary, and categorical. The toolbox also includes a suite of diagnosing methods and graphical tools to check measurements and results during and after handling missing data.

This research aims to evaluate the Type 2 Diabetes (T2D) diagnosis and prognosis power from heterogeneous environmental, lifestyle and biochemistry data. Model estimation has previously addressed three main actions as: 1) Missingvalue imputation using specific univariant and multivariant imputers accommodated to each particular feature; 2) Quasi-constancy detection in variables; 3) Constructing geographical pollution and rent data from municipality information. Next, different T2D diagnosis and prognosis models are fitted and evaluated, showing increasing performance as more specific features become available while the prediction cost rises as a consequence of requiring more specific data. Finally, four models are obtained: two of them for T2D diagnosis and the other two for T2D prognosis respectively, with performances ranging from 73.3 to 95.41 AUC-ROC. One pair of diagnosis and prognosis models were thought for a global testing that can be done in general locations by only asking general lifestyle-related questions. On the other hand, the other pair, which achieves higher performances, is thought to be applied in a clinical environment where it is easy to obtain more specific biochemistry measures.

Background and Objective: The main objective in this paper is to validate a machine-learning model trained to predict the 90-day risk of complications for patients with Acute Myeloid Leukemia using variables available at diagnosis. This is a first fundamental step towards the development of a tool that could help physicians in their therapeutic decisions. Methods: 266 patients and 36 variables form the training dataset collected by Hospital 12 de Octubre (Madrid, Spain). The external test cohort provided by Instituto de Investigación Sanitaria La Fe (Valencia, Spain) contains 162 observations. An XGBoost model was trained with one dataset and validated with the other. Additionally, the features were ranked by permutation importance and compared with the ELN 2022 risk classification by genetics at initial diagnosis. Results: The model was evaluated with the training cohort using leave-one-out cross-validation, reaching a ROC-AUC of 0.85. By setting the functioning point that maximises Youden’s index, 3 out of 4 patients with complications and 84 out of 100 in remission are correctly classified. The model was validated with external data collected in a different hospital, achieving 0.7 ROC-AUC. At the best functioning point, almost 6 out of 10 patients with complications and 8 out of 10 patients in remission are correctly classified. Ranking the variables by descending importance, the top four are, in order: age, white-blood-cells count, Gender, and TP53. The list exhibits good coherence with the ELN 2022 risk classification. Conclusions: The model achieves performances that suggest it could be used as a therapeutical decision support tool. Important variables are coherent with ELN 2022 risk classification. Further work is needed to understand the reasons for the drop in test performance. The 90-day model should be supplemented by others that predict the risk of complications at six months or one year.

Unplanned hospital readmissions mean a significant burden for health systems. Accurately estimating the patient’s readmission risk could help to optimise the discharge decision-making process by smartly ordering patients based on a severity score, thus helping to improve the usage of clinical resources. A great number of heterogeneous factors can influence the readmission risk, which makes it highly difficult to be estimated by a human agent. However, this score could be achieved with the help of AI models, acting as aiding tools for decision support systems. In this paper, we propose a machine learning classification and risk stratification approach to assess the readmission problem and provide a decision support system based on estimated patient risk scores.