Şerbetci Necip Oğuz, Blüher Stefan, Gellert Paul, Leser Ulf
Institute for Computer Science, Humboldt-Universität zu Berlin, Berlin 10099, Germany.
Institute of Medical Sociology and Rehabilitation Science, Charité, Berlin 10117, Germany.
JAMIA Open. 2025 Jul 29;8(4):ooaf064. doi: 10.1093/jamiaopen/ooaf064. eCollection 2025 Aug.
With extended life expectancy, the number of people in need of care has been growing. To optimally support them, it is important to know the patterns and conditions of their daily life that influence the need for support, and thus, the classification of the care need. In this study, we aim to utilize a large corpus consisting of care benefits applications to do an explorative analysis of factors affecting care need to support the tedious work of experts gathering reliable criteria for a care need assessment.
We compare state-of-the-art methods from explainable artificial intelligence (XAI) as means to extract such patterns from over 72 000 German care benefits applications. We train transformer models to predict assessment results as decided by a Medical Service Unit from accompanying text notes. To understand the key factors for care need assessment and its constituent modules (such as mobility and self-therapy), we apply feature attribution methods to extract the key phrases for each prediction. These local explanations are then aggregated into global insights to derive key phrases for different modules and severity of care need over the dataset.
Our experiments show that transformers-based models perform slightly better than traditional bag-of-words baselines in predicting care need. We find that the bag-of-words baseline also provides useful care-relevant phrases, whereas phrases obtained through transformer explanations better balance rare and common phrases, such as diagnoses mentioned only once, and are better in assigning the correct assessment module.
Even though XAI results can become unwieldy, they let us get an understanding of thousands of documents with no extra annotations other than existing assessment outcomes.
This work provides a systematic application and comparison of both traditional and state-of-the-art deep learning based XAI approaches to extract insights from a large corpus of text. Both traditional and deep learning approaches provide useful phrases, and we recommend using both to explore and understand large text corpora better. We will make our code available at https://github.com/oguzserbetci/explainer.
随着预期寿命的延长，需要护理的人数一直在增加。为了对他们提供最佳支持，了解影响支持需求的日常生活模式和状况以及护理需求的分类非常重要。在本研究中，我们旨在利用一个由护理福利申请组成的大型语料库，对影响护理需求的因素进行探索性分析，以支持专家收集可靠的护理需求评估标准这一繁琐工作。
我们比较了可解释人工智能（XAI）中的先进方法，作为从超过72000份德国护理福利申请中提取此类模式的手段。我们训练变压器模型，根据医疗服务单位从附带的文本注释中确定的结果来预测评估结果。为了理解护理需求评估及其组成模块（如行动能力和自我治疗）的关键因素，我们应用特征归因方法为每个预测提取关键短语。然后将这些局部解释汇总为全局见解，以得出数据集中不同模块和护理需求严重程度的关键短语。
我们的实验表明，基于变压器的模型在预测护理需求方面比传统的词袋基线表现略好。我们发现词袋基线也提供了与护理相关的有用短语，而通过变压器解释获得的短语在平衡罕见和常见短语方面表现更好，例如只提到一次的诊断，并且在分配正确的评估模块方面表现更好。
尽管XAI的结果可能变得难以处理，但它们让我们在除现有评估结果之外没有额外注释的情况下，了解数千份文档。
这项工作提供了传统和基于深度学习的最先进XAI方法的系统应用和比较，以从大量文本语料库中提取见解。传统方法和深度学习方法都提供了有用的短语，我们建议同时使用这两种方法来更好地探索和理解大型文本语料库。我们将在https://github.com/oguzserbetci/explainer上提供我们的代码。
