In these early disease diagnosis methods, rule-based research primarily relied on the analysis of medical literature and case data. Expert rule-based disease diagnosis methods involved the collection of expert diagnostic experiences to form disease diagnosis pathways, thus creating expert systems. A typical example of such an expert system is the MYCIN expert system [9], developed by Shortliffe in 1976, which became a foundational model for many subsequent expert systems in the medical field. However, many techniques are unable to explain the processes involved in disease monitoring and data inference. To address this, Aamir et al. [10] proposed a fuzzy rule-based diabetes classification system which combines fuzzy logic with the cosine amplitude method. They developed two fuzzy classifiers, and the proposed model demonstrated high predictive accuracy. Sanz et al. [11] further developed a new approach based on the Fuzzy Rule-Based Classification System (FRBCS) by integrating it with Interval-Valued Fuzzy Sets (IV-FRBCS). They validated the applicability and effectiveness of this method in medical diagnostic classification problems, demonstrating it potential in improving diagnostic accuracy in comple medical scenarios. With the accumulation of large-scale clinical data, statistical analysis has become an important method for disease diagnosis. Researchers use statistical analysis to uncover potential correlations between patient characteristics and medical indicators, thereby providing new perspectives and approaches for disease prediction and diagnosis. As a result, statistically based disease diagnosis methods have become increasingly significant in medical research and practice. Yadav et al. [12] discussed statistical modeling and prediction techniques for various infectious diseases, addressing the issue of single-variable data related to infectious diseases. They proposed fitting time series models and making predictions based on the best-fit model, offering a more accurate approach to forecasting the spread and development of infectious diseases.

Recently, neural network-based disease diagnosis methods have emerged and gradually become a hot topic of research. Researchers are focused on developing new algorithms and models that integrate multi-source information, such as clinical data, medical imaging, and biomarkers, to enhance the accuracy and reliability of diagnoses. These advancements aim to improve diagnostic capabilities by leveraging the power of deep learning in processing complex medical data. Wang and Li employed machine learning methods, combined with large-scale clinical databases, to develop a statistical-based disease prediction model. This model successfully achieved diagnostic predictions for multiple diseases, offering new insights into personalized disease diagnosis. Their approach represents a significant advancement in tailoring diagnoses to individual patient data. On the other hand, to address the challenge of assigning specific stages in the diagnosis of clinical diseases with long courses and staging characteristics, Ma et al. [13] proposed a gout staging diagnosis method based on deep reinforcement learning. They first used a candidate binary classification model library to accurately diagnose gout, and then refined the binary classification results by applying predefined medical rules for gout diagnosis. Additionally, some scholars have integrated diagnostic rules into deep neural networks for bladder cancer staging, significantly mitigating the drawback of cancer staging methods based on deep convolutional neural networks that tend to overlook clinicians’ domain knowledge and expertise. Additionally, to address the impact of imbalanced medical record samples on the training and predictive performance of disease prediction models, the academic community has proposed various solutions for training models on small and imbalanced datasets [14,15]. These methods aim to improve model robustness and accuracy when faced with limited or skewed data distributions, which is a common challenge in medical data analysis.

However, despite the numerous advantages brought by neural network-based disease diagnosis methods, there are still some shortcomings. First, the aforementioned methods often require large amounts of labeled data for training, and annotating medical data typically demands significant time and effort from expert physicians, making it costly and time-consuming. Second, due to the complexity and diversity of medical data, existing machine learning models may struggle to adapt well to various diseases and clinical scenarios, leading to insufficient generalization capability of the models. Through further reading and research, it was found that many scholars have proposed different improvement methods to address this challenge. For example, Luo et al. [16] first proposed a new multi-modal heterogeneous graph model to represent medical data, which helps address the label allocation challenges within the same cluster, enabling more precise targeting of desired medical information. Similarly, Xie et al. [17] proposed a knowledge-based dynamic prompt learning (KBDPT) method that integrates medical knowledge with pre-trained language models, enhancing diagnostic models’ expressiveness and generalizability. At the same time, Ge et al. [18]. proposed a domain knowledge-enhanced multi-label classification (DKEC) method for electronic health records, addressing the issue of previous work neglecting the incorporation of domain knowledge from medical guidelines. They introduced a label attention mechanism and a simple yet effective group training method based on label similarity. This method greatly improves applicability to minority (tail) class label distributions. Moreover, cross-modal knowledge enhancement approaches have gained attention. For instance, the DIE-CDK [19] method combines visual and textual information, leveraging cross-modal knowledge to enhance fine-grained classification, particularly for tasks involving complex features. Inspired by this approach, this study integrates medical knowledge graphs and multimodal information to improve model accuracy and adaptability in complex diagnostic tasks. Compared to existing methods, the proposed prompt-learning-based knowledge enhancement framework not only leverages cross-modal knowledge but also uses designed prompt templates to better understand structured information in medical knowledge graphs. This approach expands the model’s medical knowledge coverage, improves comprehension of complex disease characteristics, and significantly enhances model interpretability and robustness, demonstrating superior adaptability and accuracy in diagnostic tasks.

BERT (Bidirectional Encoder Representations from Transformers), proposed by Google in 2018, is an advanced pre-trained natural language processing (NLP) model. This model has achieved state-of-the-art results in various NLP tasks, including but not limited to text classification, named entity recognition (NER), and question-answering systems (QA). Recently, the research community has begun exploring the potential of the BERT model in the medical field. By integrating multi-source information such as clinical texts and medical literature, the BERT model can absorb and learn rich medical knowledge, thereby providing more precise auxiliary information in the disease diagnosis process. The bidirectional contextual modeling capability of the BERT model gives it a significant advantage in understanding the complex contexts of medical texts, offering comprehensive informational support for disease diagnosis.

In the research on medical named entity recognition algorithms, Tian et al. [7] first proposed a method based on pre-trained language models. This method utilizes the BERT model to generate sentence-level feature vector representations of short text data and combines it with recurrent neural networks (RNN) and transfer learning models to classify medical short texts. Xu et al. [20] developed a medical text classification model (CMNN) that combines BERT, convolutional neural networks (CNN), and bidirectional long short-term memory networks (BiLSTM) to address the efficiency and accuracy challenges in medical text classification. This model showed improvements in evaluation metrics such as accuracy, precision, recall, and F1 score compared to traditional deep learning models.

As advanced NLP technologies gradually penetrate the medical field, the BioBERT model has emerged. This model is specifically optimized for the biomedical domain, demonstrating significant advantages in the understanding and processing of medical texts. Sharaf et al. [21] provided a detailed analysis and overview of a systematic BioBERT fine-tuning method aimed at meeting the unique needs of the healthcare field. This approach includes annotating data for medical entity recognition and classification tasks, as well as applying specialized preprocessing techniques to handle the complexity of biomedical texts.

Recently, pre-trained language models have achieved significant success in many question-answering tasks [22,23]. However, while these models encompass a wide range of knowledge, they perform poorly in structured reasoning. Additionally, considering the previously mentioned issues, such as the low efficiency of medical data annotation, these limitations pose further challenges. This paper aims to enhance structured reasoning using knowledge graphs and leverage the BERT model’s outstanding performance in text classification tasks. The primary focus of this research is on how to effectively utilize language models and knowledge graphs for reasoning, thereby achieving the goal of significantly improving disease prediction accuracy.

Clinical texts contain many conceptual and structured forms of knowledge, such as symptoms, diagnoses, and treatment plans. Knowledge retrieval is capable of identifying and extracting key medical concepts and relationships. For example, “type 2 diabetes” is a common chronic disease, characterized primarily by persistently elevated blood glucose levels. The diagnostic criteria include fasting blood glucose and HbA1c levels. Common symptoms include excessive thirst, hunger, frequent urination, weight loss, and fatigue. Through knowledge retrieval, it can extract structured knowledge such as causes, related complications, risk factors, and more. This type of knowledge plays an important auxiliary role in disease prediction. A knowledge graph is utilized to structure various types of medical information, providing comprehensive support for disease diagnosis. The knowledge graph employed in this study includes entities such as “disease,” “symptom,” “diagnostic method,” and “treatment.” The relationship types span categories like “causes,” “used for treatment,” and “exhibits symptoms.” For instance, the entity “hyperglycemia” is connected to “type 2 diabetes” through a “causes” relationship, while “type 2 diabetes” is also associated with “kidney disease” through a “complication” relationship. By leveraging this graph structure, the model enhances logical reasoning during the diagnostic process by utilizing these relationship pathways, leading to more accurate disease predictions. The structured advantages of knowledge graphs make them highly relevant and applicable in medical tasks, particularly in complex and diverse disease scenarios. They effectively address the limitations of models solely relying on data training. Therefore, this paper utilizes the following methods to achieve knowledge retrieval from the text:

Given the clinical text x of a patient, we process x by performing tasks such as tokenization and part-of-speech tagging, resulting in a set of vocabulary x={w1,w2,…,wn}. First, we apply the named entity recognition (NER) model fNER to identify the set of entities Ex = {e1,e2,…,em}. (1)Ex←fNER(x)

Let the knowledge graph be 𝒢=(V,E), where the entity set is V={v1,v2,…,vk}. V may include entities such as “type 2 diabetes,” “hyperglycemia,” “insulin,” “weight gain,” etc. The edge set E may include relationships such as “is symptom,” “use for treatment,” “causes,” and others. Next, for each entity ei extracted from the text, we find the corresponding entity vi in the knowledge graph. The key to this matching process is determining whether ei and vi represent the same or similar concepts. This process can be implemented using a similarity function sim=(ei,vi). The entities ei and vi are considered a match when sim=(ei,vi) reaches a certain threshold θ. When the conditions ∀ei∈Ex,∃vi∈GE and ∀ei∈Ex,∃vi∈GE:sim(ei,vi)≥θ are satisfied, they are considered to represent the same or related entities.

Next, for each pair of matched entities, we search for all possible reasoning paths P in the knowledge graph. Relationships in the knowledge graph can be represented in the form of triples (vi,r,vj), where vi,vj∈G are entities, and r∈GR is the relationship. The goal of inference is to find possible paths P from entity vi to vj. (2)P(vi,vj)={(vi,r1,vi1),(vi1,r2,vi2),…,(vin,rn,vj)}

For each pair of matched entities (vi,vj), collect all possible reasoning paths, P(vi,vj), thereby obtaining structured knowledge between entities extracted from clinical texts. This knowledge is then utilized for further analysis and applications, such as disease prediction and relationship inference. For example, the entity extracted from the text et = “type 2 diabetes”, and the next step is to find the matching entity in the knowledge graph vt = “Type 2 Diabetes”. If the similarity function sim (et,vk)≥θ, they are considered a match. Finding the path P: (“Type 2 Diabetes”, “causes”, “High Blood Sugar”) → (“High Blood Sugar”, “leads to”, “Kidney Disease”). Next, collect and record the path, constructing the reasoning chain from “Type 2 Diabetes” to “Kidney Disease”. (3)Paths(vi,vj)=⋃all valid PP(vi,vj)

The following section presents a simplified example of a knowledge inference path related to “type 2 diabetes” as shown in Fig. 2. This visualization highlights the logical connections and relationships encoded in the knowledge graph, aiding the diagnostic reasoning process.

Next, we will represent the structured knowledge along these reasoning paths.

The reasoning paths from the collection are concatenated into a single text sequence. This concatenated text is then represented as a vector k using a model.

(1) Reasoning Path Representation

Each reasoning path P(gi,gj) is converted into a readable text sequence, with each path consisting of a series of triples (gi,r,gj). These triples can be transformed into text, where each triple is expressed in the form of “Entity 1 reaches Entity 2 through Relationship.” For example, given a reasoning path (g1,r1,g2), (g2,r2,g3), it can be converted into the text: “Entity g1 reaches Entity g2 through Relationship r1, and g2 reaches Entity g3 through Relationship r2.” Assuming the knowledge graph contains the following triples: (“Type 2 Diabetes”, “causes”, “High Blood Sugar”), (“High Blood Sugar”, “leads to”, “Kidney Disease”), the conversion to text would be: “Type 2 Diabetes reaches High Blood Sugar through causes, and reaches Kidney Disease through leads to.”

After converting the path into text, the pre-trained BERT model M is used to transform the text TP into a vector representation kp.

(2) Text Vectorization Model Selection (BERT)

The input text TP is first tokenized by the BERT model, breaking it down into subword units. For example, the input text “Type 2 diabetes through causes to hyperglycemia and leads to kidney disease” might be tokenized into subwords such as “Type 2,” “diabetes,” “through,” “causes,” etc. Each subword unit is mapped to its corresponding ID in the BERT vocabulary. After tokenization, the units are encoded into embedding vectors. These embedding vectors are then processed through BERT’s multi-layer Transformer encoder, capturing the contextual relationships and dependencies among the tokens. This process generates a semantic representation of the text, expressed by the following formula: (4)fBERT:xp=fBERT(TP)where:

Tp is the input text sequence.

xp is the text vector representation output by the BERT model.

The function f_BERT transforms the text sequence Tp into its vector representation xp. Here, xp contains the semantic information of the input text Tp.

The BERT model outputs the vector representation of the text Tp as: (5)kp=M(xp)

Here, xp is the vector representation obtained by processing the input text sequence Tp through the BERT model, kp represents the final text vector representation obtained through the processing of the mapping function M. The final text vector representation is obtained through a mapping function M. This additional mapping function consists of one or more fully connected layers combined with activation functions (e.g., ReLU) or nonlinear transformations. These layers can perform dimensionality reduction or expansion on the high-dimensional vectors output by BERT, aligning them with the semantic space of the knowledge graph. The function takes the BERT output embeddings as input and further processes them, adapting the general embeddings from BERT into task-specific semantic representations suitable for medical applications.

(3) Output Vector Extraction

In the text vectorization model selection mentioned earlier, we used the BERT model to convert the input text sequence Tp into the vector representation xp. During this process, we typically choose to use the special token [CLS], and its corresponding vector is used as the overall representation of the entire input text. This comprehensive representation is denoted as kp[0].

The design of prompt templates plays a critical role in this method. In this study, templates incorporating soft tokens and masked tokens (MASK) were designed to enhance the model’s understanding of medical information. For instance, for the input “Type 2 diabetes is a chronic disease,” a corresponding template might be “Type 2 diabetes is a [MASK] chronic [SOFT] disease,” where [MASK] and [SOFT] represent key positions requiring the model’s reasoning. In this template, soft tokens help the model capture the diverse characteristics of different diseases, reducing noise during the template generation process and enhancing the model’s generalization ability. Masked tokens focus the model’s attention on critical reasoning content, minimizing distractions from irrelevant information and improving diagnostic accuracy. For example, when the input text is “The patient exhibits symptoms of hyperglycemia, suspected to have Type 2 diabetes,” the template can be constructed as “The patient exhibits [MASK] symptoms, suspected to have [SOFT] disease.” This template enables the model to extract the association between “hyperglycemia” and “Type 2 diabetes,” emphasizing specific symptoms while maintaining a comprehensive consideration of disease categories. The following sections detail the steps for constructing prompt templates and configuring specific tokens to ensure the model input conforms to BERT’s optimal format, enhancing its performance in diagnostic tasks.

(1) Construction of Prompt Template

To effectively utilize the BERT model for text vectorization, we first need to preprocess the original input text x to construct a proper prompt template Template(x) that meets the input format requirements of the BERT model. This prompt template will generate a text x′ that contains special tokens and specific vocabulary. The text x′ contains special tokens such as [MASK], soft tokens, and other relevant vocabulary. These elements help the model focus on specific parts of the input and guide the prediction process, enhancing the performance of tasks like classification or inference.

For example:

Original input text x: “Type 2 diabetes is a chronic metabolic disease.”

Prompt template Template(x): “Type 2 diabetes is a [MASK] [SOFT] disease.”

Generated preprocessed text x′: “Type 2 diabetes is a [MASK] [SOFT] disease.”

(2) Expression of Template Conversion (6)x′=Template(x)=[w1,soft,w2,mask,w3,…]

Here, wi represents either words from the original text or other words added based on the context. Suppose we want to convert the original text x into the template x. This can be achieved by inserting special tokens and soft labels. A specific example is as follows:

Original input text x: “Type 2 diabetes is a chronic metabolic disease.”

Converted preprocessed text x′: “Type 2 diabetes is a [MASK] [SOFT] disease.” It can be represented as:

x′ = Template (x) = [“二型糖尿病”, soft, “是一种”, mask, “性”, soft, “疾病”.]

(3) Vector Representation and Processing

Using the pre-trained model M, the preprocessed text x′ is transformed into a set of vectors {k1,ksoft,k2,kmask,k3,…}. Each vector ki corresponds to a word or token in the preprocessed text x′, where:

ksoft is the vector corresponding to the soft token. kmask is the vector corresponding to the mask token.

For example, for the original input text x: “Type 2 diabetes is a chronic metabolic disease.”, the generated vector set can be explained as follows:

k1: The vector corresponding to the first word in the original text, “Type 2 diabetes”.

ksoft: The vector corresponding to the soft token.

k2: The vector corresponding to the second word in the original text, “is a”. kmask: The vector corresponding to the mask token.

k3: The vector corresponding to the third word in the original text, “disease”.

Other vectors follow similarly, with each vector representing the respective words or tokens in the preprocessed text.

The text x′ containing the mask and the new soft token is input into a pre-trained language model for forward inference. This produces representations for each token. The representation at the mask position is then used for a verbalizer prediction.

Given the modified input x′, it is formalized as: (7)x′=[w1,softnew,w2,mask,w3,…]

Feed x′ into the pre-trained language model M, and perform forward inference to obtain the vector representations for each token: (8){k1,ksoftnew,k2,kmask,k3,…}

Extract the representation kmask of the mask token from the model output.

Use a verbalizer function V to map kmask to a specific vocabulary term. The verbalizerV is a mapping from the vector space to the vocabulary, typically used to convert the model’s predicted vector into a specific word. (9)predicted word=V(kmask)

The specific steps are as follows:

(1) Calculate the similarity between kmask and the embedding vectors of each word in the vocabulary to obtain a probability distribution. For example, if the embedding vector for each word in the vocabulary is ei, then the probability distribution is given by: (10)similarity=kmask⋅ei

(2) Based on the computed similarities, select the word with the highest probability as the prediction result.

(3) Vector Processing

The formula for replacing the original ksoft vector with the average of ksoft and other related vectors ki is as follows: (11)ksoftnew=1n∑i=1nkiwhere:

ksoftnew is the updated ksoft vector, ksoft is the original soft token vector, ki represents related vectors (e.g., corresponding to related tokens or words), n is the number of related vectors used in the calculation.

Finally, after replacing the original ksoft with ksoftnew, we obtain the updated vector set: {k1,ksoftnew,k2,kmask,k3,…}, which will be used in the next steps of the model processing.

To better understand how the construction of prompt templates affects model performance, this study compares different template designs in subsequent experiments. For instance, besides the designed template “Type 2 diabetes is a [MASK] chronic [SOFT] disease,” we also tested a simple template without soft tokens and masked markers: “The patient exhibits symptoms, suspected to have a disease.” In this comparative template, the model directly receives complete sentences without specific markers guiding its focus on particular features and diagnostic relationships.

In this paper, experiments were conducted on the CHIP-CTC, IMCS-V2-NER, and KUAKE-QTR datasets. CHIP-CTC, one of the experimental datasets, originates from a bench-marking task released at the CHIP2019 conference. As shown in Table 1, the number of samples in the CHIP-CTC, IMCS-V2-NER, and KUAKE-QTR datasets includes the training, validation, and test samples. All text data is sourced from real clinical trials, including 22,962 entries in the training set, 7682 entries in the validation set, and 10,000 entries in the test set. The dataset is available at https://github.com/zonghui0228/chip2019task3 (accessed on 26 December 2024). Although CHIP-CTC is primarily a short text classification dataset, its data includes medical information about symptoms, medical history, and other relevant details. These attributes allow us to test the model’s performance in medical text classification tasks, showcasing its ability to provide preliminary feature extraction capabilities for disease diagnosis.

The IMCS-V2-NER dataset, used as another experimental dataset, comes from the Named Entity Recognition task in the IMCS2 dataset developed by the School of Data Science at Fudan University. It includes 2472 entries in the training set, 833 entries in the validation set, and 811 entries in the test set. The dataset is available at https://github.com/lemuria-wchen/imcs21 (accessed on 26 December 2024). This dataset primarily serves to validate the model’s capability in medical entity and term recognition, which is crucial for extracting diseases, symptoms, and treatment plans in diagnostic tasks.

The KUAKE-QTR dataset, used as another experimental dataset, includes 24, 174 entries in the training set, 2913 entries in the validation set, and 5465 entries in the test set.

The dataset is available at https:/tianchi.aliyun.com/dataset/95414 (accessed on 26 December 2024). While this dataset mainly evaluates the alignment between query terms and landing page titles, its inclusion of medical terminology and entity matches also aids in testing the model’s ability to infer relationships between symptoms and diseases. Furthermore, although these datasets are not explicitly designed for disease diagnosis, they play a vital role in verifying the proposed model’s adaptability to multiple tasks and its performance in broader medical text processing scenarios. These datasets help validate the model’s effectiveness in extracting medical text information, laying the groundwork for extending its application to more complex real-world diagnostic tasks, such as clinical case reports and electronic health records.

When evaluating our proposed method, we established a comprehensive set of baseline models to ensure rigorous and fair comparison. These baselines were selected to represent robust benchmarks in the field, including both traditional machine learning algorithms and advanced deep learning techniques.

SVM [24]: Support Vector Machines (SVMs) classify text categories using hyperplanes, excelling at handling linear and nonlinear problems. SVMs are well-suited for tasks like sentiment analysis and spam detection.

CNN [25]: Convolutional Neural Networks (CNNs) extract local features from text through convolution layers, enabling quick identification of critical information. They are ideal for content understanding and classification tasks.

RNN [26]: Recurrent Neural Networks (RNNs) are particularly effective in text classification tasks due to their ability to process sequential data and capture long-term dependencies.

Attention [27]: The Attention mechanism assigns higher weights to important words or phrases, improving focus on critical information and optimizing classification performance.

BiLSTM [28]: Bidirectional Long Short-Term Memory Networks (BiLSTMs) combine forward and backward LSTM layers, offering advantages in capturing semantic dependencies within text.

BiRNN [28]: Bidirectional Recurrent Neural Networks (BiRNNs) capture information in both forward and backward text sequences, enhancing the understanding of long-term dependencies.

BERT [29]: BERT employs bidirectional encoding and pre-training to capture semantic and contextual features of text. It performs exceptionally well across various text classification tasks.

BioBERT [30]: BioBERT is a pre-trained model specifically designed for biomedical text. Trained on medical literature, it excels in understanding medical terminology and performing biomedical tasks. Compared to general-purpose BERT, BioBERT is better suited for processing medical and biological contexts.

These baseline models provide a reliable reference for evaluating our approach, enabling a clearer demonstration of its advantages in addressing specific text classification tasks.

To ensure the reliability of the experiments, we repeated each experiment three times and used the average values as the final results. The experimental results are shown in Table 2 and Figs. 3–6.

This section aims to analyze the effectiveness of each module in our proposed method. To achieve this, we first remove the knowledge representation component from our method. As seen in the first row of Table 3, removing the knowledge representation results in a significant drop of 0.2 in the F1 score. This is because knowledge representation is a core component of our method, which injects structured knowledge obtained from the knowledge graph into the language model, thereby enhancing the model’s understanding of medical domain knowledge and improving performance on medical diagnostic tasks.

Subsequently, we modified the prompt template, changing it to: “The characteristics of Type 2 Diabetes include [MASK] nature, and it is also manifested as [SOFT] disease.” From rows 2 to 4 of Table 3, it can be observed that the prompt template used in our method achieves the best performance, while other prompt templates do not perform as well. We believe this may be due to the presence of excessive or irrelevant words in the prompt templates, which could introduce noise into the language model’s reasoning process.

This study conducts systematic ablation tests to evaluate the impact of knowledge enhancement modules, diverse prompt template designs, and label verbalization strategies on disease diagnosis performance. It compares the efficacy of basic symptom templates, explicit symptom-disease relation templates, and complex context templates in contextual encoding, while assessing label mapping approaches including direct term matching, expanded vocabulary, and synonym tolerance. Findings highlight that knowledge integration strengthens reasoning capabilities, whereas excessive contextual complexity introduces noise.

In this study, we designed various templates and verbalizers to enhance the model’s performance in medical text classification tasks. Basic Symptom Template: Focuses on symptom information for inference, suitable for simple symptom classification tasks. Symptom-Disease-Relation Template: Introduces symptom-disease associations, improving the model’s diagnostic accuracy. Complex Context Template: Handles more intricate contextual relationships, such as patient medical history or symptom combinations, enhancing the model’s performance in multidimensional scenarios. On Verbalizers, Direct Mapping Verbalizer: Simplifies reasoning by directly mapping symptoms to specific diseases. Extended Vocabulary Verbalizer: Expands the vocabulary to improve recognition of different expressions and synonyms. Synonym Tolerant Verbalizer: Strengthens the model’s understanding of synonyms and semantically similar terms, improving robustness across diverse expressions.

The combination of these templates and verbalizers allows the model to flexibly handle different layers of information in medical text, achieving favorable results in disease diagnosis tasks.

This section aims to analyze the interpretability of the model. Unlike general text classification tasks, in disease diagnosis tasks, users are additionally concerned with the interpretability of the results. An unexplainable prediction result is unacceptable to users. Therefore, to verify the model’s explainability across diverse clinical cases, we introduced a multi-case validation approach to ensure consistency and transparency in its reasoning processes. The cases include diabetes, hypertension, hyperlipidemia, and coronary heart disease, covering a variety of symptoms, medical histories, and complex comorbidity scenarios. This ensures the comprehensiveness and applicability of the explainability analysis. For example, a case involving a 55-year-old male patient with diabetes was selected to visualize the reasoning process through illustrative diagrams.

Using the clinical case in Fig. 7 as an example, we can observe the knowledge retrieved during the knowledge retrieval phase, including: “Type 2 diabetes is a common chronic disease,” “Its main characteristic is persistently elevated blood sugar levels,” and “Frequent drinking, eating, and urination are common symptoms.” From this knowledge, we can identify which key pieces of information from the clinical case are related to the final prediction. In other words, we can clearly understand which knowledge elements led the model to make the current prediction.

Through analyzing multiple clinical cases, we observed that the model consistently traced its reasoning paths across different conditions, ensuring transparency and reliability in the diagnostic process. This multi-case validation approach not only improves the model’s explainability in complex clinical scenarios but also provides clinicians with clearer reasoning references, thereby enhancing the model’s credibility.

In addition, to facilitate readers’ understanding, this paper not only explains the evaluation metrics but also demonstrates the advantages of the proposed method through relevant examples. Figs. 8–10 show the specific example.

The specific examples presented show that the model proposed in this paper addresses the issues of insufficient domain knowledge and difficulties in long-tail entity recognition in existing models. The predicted disease conditions are closer to actual clinical outcomes. These examples further illustrate the innovation and superiority of our approach, allowing readers to intuitively understand the improvements and advantages over traditional methods.