Preventive medicine aims to detect potential health risks early, thereby reducing morbidity and healthcare expenses. Beyond the mere anticipation of possible diagnoses, there is a demand for systems that justify predictions with clinical evidence and allow a quantitative assessment of the quality of the explanation. In this work, we present a novel approach: first, to address the prognosis of subsequent diagnoses in Spanish clinical data and, second, to generate explanations for each estimated diagnosis using generative models. By structuring patient data with the ICD-10 coding standard, and with the aid of cross-lingual data augmentation, substantial prognosis estimation improvements were achieved. We also put special attention to the assessment of the quality of the explanations generated. In this line, we propose the FIR framework (i.e. faithfulness, interpretability, robustness), in an attempt to seize the ability of each explanation to address core clinical questions relevant to the estimated ICD-10 diagnoses. Combining ICD-10 timelines coming from Osa (Spanish EHRs dataset) with MIMIC-IV (English EHRs dataset) significantly improved our Disease Risk Identifier’s accuracy (by 11

Early detection of mental health risks on digital platforms is essential for enabling timely interventions. However, existing metrics such as Early Risk De-tection Error (ERDE) assess detection latency by message count rather than real time, misaligning with clinical priorities that depend on the temporal dynamics of risk. We propose Time-Aware Early Risk Detection Error (T-ERDE), a metric that replaces message-count penalties with temporal cost functions based on days elap-sed since the first user interaction. T-ERDE preserves compatibility with existing metrics while offering a more clinically meaningful and equitable evaluation. Expe-riments on the mentalRiskES (Spanish) and eRisk (English) datasets demonstrate that T-ERDE provides more accurate and relevant assessments, better capturing intervention urgency and reducing biases linked to user activity volume.

Objective: In the context of assessing suicide risk using questionnaires as measurement instruments, the main goals are: (i) to compare the performance of the classification task with knowledge-based algorithms and inferred approaches, and (ii) to reduce the set of questionnaires. Methods: A classification task is performed on the set of questionnaires considering two methods: expert knowledge translated to algorithms and represented as diagrams, and data-inferred machine learning models. Feature ablation is performed to reduce the questionnaire items. Results: Machine learning models are able to detect risk with an F1 macro average score of up to 85%, significantly better than knowledge-based models. The number of questionnaire items can be reduced with no significant impact. Conclusions: Inferred models can be used to predict the level of suicide risk with a reduced set of questionnaires. Moreover, the Suicide Cognitions Scale-Revised questionnaire is seen to be the most impactful in the prediction.

The optimization of early mortality prediction models to accurately identify high-risk patients during a hospital admission is key in medical research to assist in healthcare decision-making. While these models use structured (physiological features) or unstructured (free-text clinical notes) data, integrating them synergistically, especially for early prediction, remains a challenge. We introduce a novel, model-agnostic approach to represent clinical notes as quantitative, task-specific time-series, structurally aligning them with physiological features. Instead of relying on complex multimodal architectures, we shift fusion from the architecture level to the data level by transforming each note into a numerical mortality risk score, creating a new time-series feature that can be seamlessly integrated into any standard time-series prediction model. We evaluate our approach on the following: first, early prediction, second, data scalability, and third, versatility across centralized, federated, and local learning frameworks, in an attempt to assess the stability of the approach in scenarios with different amounts of data to learn from. Using only readily accessible data from the first 8-24 h postadmission, our method significantly boosts baseline model performance, with relative improvements of up to 7 % in area under the receiver operating characteristic curve and 28 % in area under the precision-recall curve over prior work on early prediction that relies solely on physiological data. Moreover, we find that incorporating our text-derived scores significantly achieves higher performance gains than nearly doubling the training data size in scenarios where no textual information is considered. Rather than proposing a predictive model, we present a new technique to structure clinical narratives, enabling a more effective fusion with temporal, quantitative health data for improved early prediction of mortality.

Predicting future diagnostic codes from Electronic Health Records (EHRs) is a complex task due to the unstructured nature of medical texts. This study introduces RiskProg, a system designed to predict future diagnoses from Spanish EHRs. It aims to structure EHRs through generative LLMs and then predict future diagnoses by highlighting the most influential words that contributed to the prediction. RiskProg utilizes a three-step approach: (1) a summarizer that structures patient visit information into comprehensive summaries, (2) a future potential diagnoses predictor, and (3) a salient word identifier based on integrated gradients to highlight the textual elements influencing predictions. The system was evaluated using the Osa dataset, which includes 26,731 annotated Spanish discharge summaries. The incorporation of summarized EHR documents enhanced classification accuracy, with expanded summaries outperforming the baseline classification with raw documents. Explainability, achieved through word-level attributions, increased clinician trust by providing insights into the model's decision-making process. RiskProg showcases the potential of structuring unstructured clinical data for next diagnosis prediction. The approach offers both an explanation of the predicted codes and structured summaries that make it easier to assess patient admission information.

A bottleneck in the Electronic Health Records (EHRs) classification according to the International Classification of Diseases (ICD) task is the challenge involved in getting large amounts of clinical Spanish documents for training efficient language models with private health data. The federated learning (FL) strategy enables the independent training of several models and the subsequent unification of each resulting model parameters to generate a unified model without the need to share sensitive data out of the clinical facilities. We analyse the feasibility of employing the federation strategy in Spanish in the context of an actual data division environment: data coming from two real hospitals from the Basque health system and generated in the same period. We also propose a method to further pre-train the language model (LM) in a federated manner. We apply our federated further pre-training method to the training of BETO and BERTmultilingual. Our findings clearly show that it is feasible to carry out federated learning for Spanish EHR classification using data spread across different hospitals. Moreover, the proposed LM further pre-training method steadily surpasses the results of the model without further pre-training.

Sare sozialak norbere esperientziak adierazteko tresna garrantzitsu bihurtu dira. Esperientzia negatiboak kontatzeko ihesbide ere badira, baita ere gaitz mentalak dituzten pertsonentzat. Lan honek, hain zuzen, sare sozialetan suizidio-zantzuak erakusten dituzten pertsonak ditu jomuga, euren testuetan ezaugarri linguistiko bereizgarriak bilatzen aritu baikara, halako kasuen atzemate automatikoa errazte aldera. Hartarako, euren mezuak oinarritzat hartu eta sistema automatikoen bidez azterketa morfosintaktikoa eta sentimenduen analisia egin dugu. Lexikoa kuantitatiboki alderatu dugu ertz ugaritatik, tartean sistema adimendunak erabiliz. Suizidio-zantzuak adierazten dituzten pertsonek idazterakoan ezaugarri morfosintaktiko bereizgarriak dituztela berresteaz gain, erabiltzaile horiek sentimendu positibo zein negatiboak gai pertsonaletara bideratzeko joera nabarmena dutela ondorioztatu dugu.

Background and Objective: Healthcare documentation processing is becoming more and more efficient and effective as a result of advances in machine learning and natural language processing (NLP). One challenge in clinical practice is the early detection of future patient potential diagnoses, which is crucial for preventive medicine. Estimating the potential future diagnoses, helps to speed up the management of Electronic Health Records (EHRs) and opens a path towards clinical prevention. It is a challenging task, as there are thousands of possible diseases, and, in general, there is limited data available to train systems due to privacy concerns.The objective of his study is to infer future probable diagnoses given patients diagnosis history. In previous works, this task has been carried out using structured data, such as, ICD-coded diagnoses, overlooking unstructured textual information in EHRs. Unlike traditional methods, this study aims to enhance next-diagnosis prediction by integrating patient diagnosis information codified according to the International Classification of Diseases (ICD) with unstructured clinical text. Methods:: We propose a multi-faceted model that integrates structured ICD-encoded patient histories with unstructured EHR text for future diagnosis prediction. Our approach consists of (1) a sequential model trained on structured diagnosis timelines, (2) a Clinical Longformer-based model trained on unstructured EHRs, and (3) an ensemble strategy to combine predictions from both components. Results:: Our proposed ensemble strategy significantly outperforms current state-of-the-art approaches in predicting future diagnoses, achieving a Precision@5 of 72.34% and a Precision@20 of 77.49%. Additionally, it showed high robustness and reliability across different demographic groups and a varying scope of medical history. Conclusion:: This research demonstrates that the integration of structured ICD diagnoses timelines with unstructured EHRs achieves improved results compared to just using structured diagnosis timelines. Notably, the proposed model also maintained high accuracy even with a short-term history of diagnoses.

The task to generate a timeline representation of the International Classification of Disease (ICD) codes included in a given Electronic Health Record (EHR) is an open challenge. One step in that direction is the identification of temporal entities and relations in the text to arrange, chronologically, the important clinical events in the document. In this work, we cope with the recognition of temporal entities and the extraction of temporal relations between them in EHRs. Current approaches usually train separate models for entity recognition and relation extraction, easing the models at the expense of ignoring the underlying dependencies. This work explores SpERT, a joint entity recognition and relation extraction approach based on the Transformer architecture. We trained and evaluated the model with a small data-set with annotated temporal information. While previous works do not test their model on other data-sets, we also apply the model to a more extensive dataset without temporal information labelling in order to evaluate the performance of the model qualitatively. The qualitative evaluation allows the discovery and interpretation of the strengths and weaknesses of the model. Our experiments conclude that SpERT results are close to the state of the art. Moreover, we discovered that the model could generalize when applied to other clinical datasets.

Background and Objective: In the realm of automatic Electronic Health Records (EHR) classification according to the International Classification of Diseases (ICD) there is a notable gap of non-black box approaches and more in Spanish, which is also frequently ignored in clinical language classification. An additional gap in explainability pertains to the lack of standardized metrics for evaluating the degree of explainability offered by distinct techniques. Methods: We address the classification of Spanish electronic health records, using methods to explain the predictions and improve the decision support level. We also propose Leberage a novel metric to quantify the decision support level of the explainable predictions.We aim to assess the explanatory ability derived from three model-independent methods based on different theoretical frameworks: SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and Integrated Gradients (IG). We develop a system based on longformers that can process long documents and then use the explainability methods to extract the relevant segments of text in the EHR that motivated each ICD. We then measure the outcome of the different explainability methods by implementing a novel metric. Results: Our results beat those that carry out the same task by 7%. In terms of explainability degree LIME appears as a stronger technique compared to IG and SHAP. Discussion: Our research reveals that the explored techniques are useful for explaining the output of black box models as the longformer. In addition, the proposed metric emerges as a good choice to quantify the contribution of explainability techniques.

The aim of Task 2 (Early Detection of Signs of Pathological Gambling) from the CLEF 2023 eRisk Work-shop is to analyze social media users’ messages for early warning signs of pathological gambling. Given that Pathological Gamblers are a small set compared to the Control group, we propose the utilization of a neural network incorporating a customized loss function to effectively tackle the challenge of class imbalance. In our proposed loss function it is possible to adjust the penalty for false positives and false negatives, increasing the penalty for the critical false negatives. Our proposed solution demonstrates robustness, achieving one of the highest recall rates while maintaining a competitive precision. Furthermore, our system introduces a range of potential variations that warrant further investigation.

Civil registration and vital statistics systems capture birth and death events to compile vital statistics and to provide legal rights to citizens. Vital statistics are a key factor in promoting public health policies and the health of the population. Medical certification of cause of death is the preferred source of cause of death information. However, two thirds of all deaths worldwide are not captured in routine mortality information systems and their cause of death is unknown. Verbal autopsy is an interim solution for estimating the cause of death distribution at the population level in the absence of medical certification. A Verbal Autopsy (VA) consists of an interview with the relative or the caregiver of the deceased. The VA includes both Closed Questions (CQs) with structured answer options, and an Open Response (OR) consisting of a free narrative of the events expressed in natural language and without any pre-determined structure. There are a number of automated systems to analyze the CQs to obtain cause specific mortality fractions with limited performance. We hypothesize that the incorporation of the text provided by the OR might convey relevant information to discern the CoD. The experimental layout compares existing Computer Coding Verbal Autopsy methods such as Tariff 2.0 with other approaches well suited to the processing of structured inputs as is the case of the CQs. Next, alternative approaches based on language models are employed to analyze the OR. Finally, we propose a new method with a bi-modal input that combines the CQs and the OR. Empirical results corroborated that the CoD prediction capability of the Tariff 2.0 algorithm is outperformed by our method taking into account the valuable information conveyed by the OR. As an added value, with this work we made available the software to enable the reproducibility of the results attained with a version implemented in R to make the comparison with Tariff 2.0 evident.