Intelligent MDT treatment decision making for stage III NSCLC using dual level embedding and three level explanation

Abstract

Stage III non-small cell lung cancer (NSCLC) presents complex challenges in treatment decisions due to extensive disease characteristics and patient heterogeneity. Multidisciplinary teams (MDT) offer personalized treatments by integrating diverse expertise. However, resource and time constraints limit MDT practicality in healthcare. Addressing this, we propose an interpretable approach for intelligent MDT treatment recommendations. Our approach focuses on enhancing clinical text representation and the interpretability of recommendations. Using a dual-level embedding technique, the local and global textual information can be captured. By delving into intricate factors at word, phrase, and sentence levels, we explain treatment recommendations for heterogeneous patients. Specially, attention flow is employed to consider the relationship between attentions across multiple layers, and attention mechanisms screen out important words and sentences. Our method achieves over 85% across accuracy, precision, recall, and F1 score in stage III NSCLC treatment recommendations. Furthermore, we assess the association between our proposed model and patient survival outcomes by analyzing patients who did not undergo MDT consultation. The results demonstrate that patients receiving model-concordant treatments exhibited significantly higher survival rates at 1, 3, and 5 years compared to those receiving model-nonconcordant treatments. Moreover, Kaplan-Meier survival curves confirm the improvement in survival outcomes associated with model-concordant treatments. Additionally, ablation analysis validates the rationality of the model structure.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-39658-2.

Introduction

Lung cancer has a high incidence and mortality rate on a global scale, with China particularly experiencing a significant impact, ranking at the forefront^1,2. Non-small cell lung cancer (NSCLC) is a common type of lung cancer, and nearly 30% of NSCLC patients have already been in stage III at the time of initial diagnosis³. However, stage III NSCLC is a heterogeneous disease with poor prognosis, for which treatment options are difficult to define^4,5. Therefore, this paper contributes to treatment decision-making for stage III NSCLC.

Multidisciplinary team (MDT) provides the most appropriate and personalized treatment for stage III NCSCL patients by integrating treatment opinions from multidisciplinary experts^6,7. Moreover, more and more studies have shown that MDT could significantly improve patient prognosis, such as improving quality of life and prolonging survival^7,8. Unfortunately, the worldwide medical resources are commonly scarce and unevenly distributed, leading to a limited number of hospitals equipped with MDTs⁹. Particularly, when facing with a large patient population, it is more difficult to ensure that the vast majority of patients have access to treatment recommendations provided by MDTs¹⁰. Although some studies provide treatment recommendations based on clinical practice guidelines^11–14 and research literature^15,16, Lin, et al. ¹⁷ pointed out that clinical practice guidelines do not respond well to patient heterogeneity and cannot provide the same personalized treatment plans as those provided by MDT. Treatment recommendations based on research literature also not take into account individual patient differences¹⁷, and some information may be out of date¹⁸, and there may be conflicting evidence affecting treatment choice¹⁹. Therefore, personalized treatment plans that consider individual patient characteristics are necessary, and there is a need to explore intelligent treatment decision-making method to provide stage III NCSCL patients with the treatment plan comparable to the one developed by MDT.

Electronic medical records (EMRs) detailing individual patient characteristics²⁰ have been used for treatment recommendations to provide personalized care^4,11,21–29. However, the information available in EMRs has not been fully utilized in these studies. Hendriks, et al. ³⁰, Lin, et al. ¹⁷, Sesen, et al. ⁴ and Yan, et al. ³¹ only selected a few factors based on experience and knowledge, Najafabadipour, et al. ²⁸, Solarte-Pabón, et al. ²⁹, Zeng, et al. ²³ and Zhang, et al. ³² only considered certain medical entities, Min, et al. ³³ and Zeng, et al. ²² only analyzed word vector information in clinical text. Furthermore, the performance evaluation of the model in current studies is primarily focused on prediction effects without analysis of the association between the recommended regimen and patient survival outcomes^22,30,31. Finally, for the sake of practical use in clinical practice, it is crucial to improve the interpretability of the model so that clinicians could understand the reasoning behind the recommended treatment and make informed decisions³¹.

To overcome the shortcomings mentioned above and to provide stage III NSCLC patients with the treatment options that closely approximate those developed by MDT, this paper puts forward an interpretable intelligent treatment decision-making method. The contributions of this paper are summarized as follows.

1. Dual-level representation of EMRs using word and sentence embeddings. The dual-level representation considers both local information representation through word embedding and global information representation through sentence embedding³², which ensures adequate representation of clinical text.

2. Treatment decision of each highly heterogeneous patient is deeply explained at three levels with the help of attention mechanisms and attention flows: word, phrase, and sentence. This three-level explanation can achieve global interpretation by considering the relationship between attentions across multiple layers, and ensure that interpretation is complete and reliable by complementing and supporting each other with explanations at three levels.

3. The proposed model is evaluated in terms of prediction performance and survival analysis. The prediction performance is assessed by analyzing the accuracy of treatment recommendation. Associations between recommended treatment and patient survival are assessed not only through 1-year, 3-year, and 5-years survival rate comparisons but also using Kaplan–Meier curves and multivariable Cox regression³³.

4. An interpretable intelligent MDT treatment decision-making method based on EMRs is proposed, which can approximate MDT recommendations with high accuracy for cancer patients in the studied dataset.

The rest of the paper is organized as follows. In Sect. 2, we review previous work on the treatment recommendation and MDT treatment recommendation. In Sect. 3, we describe the four main components of the MDT treatment decision-making model: data collection and preprocessing, dual-level embedding representation, model performance evaluation and three-level explanation. Subsequently, the prediction performance and the association between the recommended regimen and patient survival outcomes under two tertiary hospitals are analyzed, the treatment recommendations are interpreted at three levels, and the rationality and superiority of the model is verified by ablation experiments. In the final section, we summarize the findings and limitations of the study, as well as future research directions.

Prior work

MDT treatment recommendation

MDT can effectively improve patient prognosis³⁴, and enhance patient satisfaction^8,35, but at the same time MDT consumes a lot of medical resources and is very time-consuming³⁶. Therefore, some scholars have been more inclined to exploring intelligent MDT treatment recommendation.

Table 1 presents information about representative MDT treatment recommendation studies. The studies were identified through searches in Web of Science, PubMed, and Google Scholar using keywords such as “multidisciplinary team”, “treatment recommendation” and “machine learning”. Several characteristics of the current study can be found in the Table (1) Current MDT treatment recommendation studies concentrate on cancer, particularly breast cancer^21,37–39. There is a research gap in MDT treatment recommendations for lung cancer in specific stage. (2) Most studies are based on clinical practice guidelines or experience or knowledge to determine treatment recommendation variables^4,21,37–39. It is questionable whether these variables could truly and adequately reflect the individual characteristics of the patient. (3) Common interpretable machine learning methods such as decision trees, logistic regression and Bayesian networks are widely used for MDT treatment recommendations^{4,21,37,39–41}. This mainly because clinicians need to understand the underlying reasoning behind a model’s predictions in order to make informed decisions about patient care. (4) Classification model assessment metrics are widely used in MDT treatment recommendation^21,37–41, but less research has been conducted to assess the association between the recommended regimen and the prognostic effect of the recommended regimen⁴. Thus, there is still much room for development of MDT treatment recommendation research.

Table 1.

MDT treatment recommendation studies.

Study	Sample size	Disease	Factors	Method	Metric	Interpretability
30	504	Breast cancer	The factors in the Dutch breast cancer guideline	Clinical decision trees (CDT)	Concordance of recommendations generated by the MDT versus the CDTs	Explainable decision trees
31	1535	Breast cancer	12 categorical attributes	K-nearest neighbor classification (k-NN)	Accumulative accuracy	Not consider explanation
40	304	Basal cell carcinoma	6 variables after screening based on the variance-inflation factor	Classification tree	Accuracy	Explainable classification trees
4	4020	Lung cancer	13 patient and disease specific variables	Rule-based decision support approach, Bayesian networks	Confusion matrix, AUC, 1-year survival	Explainable methods
17	1924	Breast cancer	Dozens of variables	10 supervised machine learning classifiers	AUC, sensitivity, specificity, and positive likelihood ratio	Some methods are interpretable
41	66	Laryngeal cancer	303 variables	Inference algorithms and probabilistic graphical models on Bayesian networks	Accuracy, confusion matrix, ROC curve, AUC, and calibration curve	Explainable method
39	3340	Breast cancer	48 variables	Combine k-NN classifiers and rule-based methods	Hit rate (HR) and the binary classification error rate (BER)	Not consider explanation

Note: Receiver operating characteristic (ROC) curve, area under the ROC curve (AUC).

Treatment recommendation

There is a wealth of research on treatment recommendations. Table 2 summarizes representative studies on treatment recommendations published in recent years. They were selected from Web of Science, PubMed, and Google Scholar using keywords such as “treatment recommendation” and “machine learning”. The studies reflect variation in disease domains, data sources, methodological strategies, evaluation metrics and interpretability. (1) Scholars have analyzed the treatment recommendation for chronic diseases as well as some common diseases^24–27, in addition to the rich research on the treatment of cancer^{12–16,22,23,42}. (2) Not only are clinical practice guidelines and research literature used to determine treatment decision options^22–27,42, but information in the EMRs, such as clinical note, demographics, and laboratory test results, is explored to adequately represent patient characteristics in order to provide more appropriate treatment options^12–16. Moreover, information in clinical notes is often represented through medical entities or word embeddings^22,23. (3) With the development of machine learning techniques, scholars are increasingly resorting to reinforcement learning to enable treatment recommendations^24,25,27. Nevertheless, the computational complexity of reinforcement learning is high and the definition of reward functions is very difficult in medical field⁴³. (4) In addition to classification metrics^{12–16,22,26}, scholars are beginning to focus on the association between patient survival outcomes and the recommended regimen provided by treatment recommendation models^{23–25,27,42}. (5) More and more models are considering interpretability given the need for clinical practice ^{12–15,22−24,26}.

Table 2.

Treatment recommendation studies.

Study	Disease	Factors	Method	Metric	Interpretability
22	Prostate Cancer, Oropharynx Cancer, Esophagus Cancer	Clinical notes represented by ‘structured + bow’ or ‘structured + doc2vec’ or ‘structured + fasttext’	Five machine learning methods	Sensitive, Precision, F1-score	t-SNE
23	Prostate cancer, Stage I NSCLC	Clinical text represented by the TF-IDF of biomedical entities	Lasso	Hazard ratios	Explainable method
15	Lung cancer	Nine key clinical concepts	Iterative algorithms for concept recognition and recommended guidelines for treatment plans	F1-score, recall, precision	Each recommendation is supported by literature or consensus guide
42	NSCLC	127 features, including patient characteristics, tumor stage, and treatment strategies	3-layer neural network	3-year and 5-year survival	Not consider explanation
26	Low back pain	A multivariate sequence of health measurements over time (i.e., symptoms such as pain or activity limitation)	Variable-duration copula hidden Markov model	F1-score, balanced accuracy, cost, effectiveness	Three states with clinically relevant interpretations
27	Type 2 Diabetes	Demographic information, physical measurements, medical history, and laboratory data	Reinforcement learning	Short-term and long-term outcome	Not consider explanation
14	Colorectal cancer	11 concept groups	Guideline-based treatment evaluation	Precision, recall, accuracy, and F1-score	Explainable method
25	Top 2,000 diseases in Multiparameter intelligent monitoring in intensive care database	Demographics, lab values, vital signs, and output events	Supervised reinforcement learning with recurrent neural network	The mortality rates, Jaccard coefficient	Not consider explanation
24	Comorbidity and Sepsis	Demographics, lab values, vital signs	Hierarchical imitation learning	Mortality rate, AUC and Jaccard coefficient	Providing explainable recommendation with hierarchical structure
16	Various cancers	COSMIC Cancer Gene Census, COSMIC Mutation Data, Genomic Data Commons and 26 million Medline abstracts	Generative adversarial neural network	Precision, recall, infNDCG, R-prec, P@5, P@10, P@30, S@10	Not consider explanation
13	Breast cancer	Guideline, domain knowledge	A care plan builder based on rules	Accuracy	Explainable method
12	Breast cancer	Three Clinical practice guidelines (NCCN、AP-HP and ONK)	The guideline-based decision support module	The compliance rate of guidelines with the recommended treatment	Explainable method

A review of the literature listed in Tables 1 and 2 reveals the following three points for improvement. First of all, most studies rarely consider and even ignore MDT treatment decisions. To put it another way, they focused on the use of clinical practice guidelines or research literature or expert experience to develop treatment regimens with strong generalizability and weak personalization. Unluckily, patients with stage III NSCLC are highly heterogeneous and require adequate analysis of individual patient characteristics to provide personalized treatment plans. Secondly, the existing literature mainly assesses prediction outcomes through indicators such as accuracy and F1-score, while the association between the recommended regimen and patient survival outcomes has received relatively little attention, despite its relevance for assessing the clinical utility of treatment recommendations. The model should not only have good predictive effect, but also bring practical application value. Third, most existing studies tend to use word embeddings or medical entities to represent clinical texts, which only reflects local information in the text and cause information loss to a certain extent. Finally, from the perspective of model interpretability, previous studies gradually highlight the transparency and interpretability of models. However, there is still a need to continue to improve the interpretability of the models to promote their practical application.

Methodology

In this paper, we built an interpretable MDT treatment decision-making model consisting of four components. (1) Data collection and preprocessing: Collect EMRs of patients with stage III NSCLS, determine the treatment for recommendation and divide the patients into MDT and non-MDT groups. (2) Dual-level embedding representation: word and sentence embedding representation. (3) Model performance evaluation: assessment of prediction performance and evaluation of associations with survival outcomes. (4) Three-level explanation: word, phrase, and sentence levels model interpretation. The framework of the proposed model is shown in Fig. 1.

Fig. 1.

The framework of the proposed model.

Data collection and preprocessing

We collected EMRs of patients with stage III NSCLC between January 1, 2017 and October 30, 2022 from two tertiary hospitals in Hunan Province, China. Patients with the first diagnosis of stage III NSCLC and patients diagnosed with stage III NSCLC since recurrence or progression were all the research sample. Samples for which no treatment plan given were excluded. Ultimately, a total of 2,876 stage III NSCLC patients were used for this study. The present study was approved by the Ethics Committee of The Second Hospital of Central South University and Hunan Cancer Hospital, and conducted in accordance with the principles of Declaration of Helsinki. All patients signed written informed consent.

There are various types of notes in the EMRs. From the 36,490 collected notes for these patients, progress notes are selected as the basis for recommending treatments and treatment plan sheets are chosen to outline the specific treatment plans for these patients. To prevent potential information leakage, only progress notes documented from the time of diagnosis up to the date of the MDT discussion or treatment decision were included. Notes written after treatment initiation were excluded. In addition, explicit future-oriented or order-set expressions were manually screened and removed to ensure that the input text did not directly reveal the target treatment. This manual screening was independently performed by three trained researchers using a predefined set of treatment-related keywords. Representative keywords included “start concurrent chemoradiation”, “initiate cisplatin”, “schedule radiotherapy”, “begin adjuvant chemotherapy”, “prescribe paclitaxel”, and “administer immunotherapy” among others related to treatment initiation, drug orders, and procedural plans. The retained progress notes mainly contained clinical and diagnostic information, such as positron emission tomography-computed tomography (PET-CT), lung-enhanced computed tomography (CT), biopsy, immunohistochemistry, magnetic resonance imaging of the head, abdominal CT, etc., as well as demographic information of the patients, such as age, gender, clinical stage, etc. This information gives a full picture of the patient’s basic profile and provides the basis for the doctor’s decisions. A representative example of such a progress note, highlighting the clinical and diagnostic information used for treatment recommendation, is provided in Supplementary Material.

On the basis of the treatment plan sheets in EMR, we continued to consult clinical experts, studied clinical practice guidelines and a large amount of literature^4,44. Following this, the collected samples were divided the treatment into the following six groups according to the specific situation of them. The 6 treatment plans are surgery, surgery after neoadjuvant therapy, radiotherapy, chemotherapy, chemoradiotherapy (sequential/concurrent) and chemotherapy + immunotherapy/targeted therapy, respectively. Treatments that only involve a very small number of patients are not considered. Finally, a total of 2,521 patients who chose one of these six treatment options were included in the investigation. Among them, 22.8% of patients received a consultation from the MDT. The proportion of patients who did and did not receive MDT consultation (non-MDT) under each treatment option is shown in Table 3.

Table 3.

Treatments received by patients.

MDT	SUR	SURANT	CHE	RAD	CHE (S/C)	CHE + IMM/TAR
	23.69%	21.25%	8.01%	5.92%	19.86%	22.12%
non-MDT	17.57%	3.65%	36.67%	1.59%	8.58%	31.84%

Note: SUR refers to ‘Surgery’, SURANT refers to ‘Surgery after neoadjuvant therapy’, CHE refers to ‘Chemotherapy’, RAD refers to ‘Radiotherapy’, CHE (S/C) refers to ‘Chemoradiotherapy (Sequential/Concurrent)’, CHE + IMM/TAR refers to ‘Chemotherapy + immunotherapy/targeted therapy’.

As can be seen in Table 3, the two patient groups (MDT vs. non-MDT) used different treatment protocols. The MDT group made greater attempts at various treatment options, whereas patients in the non-MDT group were more inclined to using chemotherapy + immunotherapy/targeted therapy. Compared to the non-MDT group, the MDT group had a higher proportion of surgery after neoadjuvant therapy, reduced the proportion of chemotherapy and increased the proportion of chemoradiotherapy.

The observed differences in treatment patterns between the MDT and non-MDT groups likely reflect the advantages of multidisciplinary review in NSCLC management. Multidisciplinary experts can better cope with the heterogeneity of NSCLC patients, comprehensively analyze the specific conditions of the patients, and provide them with targeted treatment options. While MDTs may indeed manage a higher proportion of complex cases, these differences also suggest that MDTs are able expand therapeutic options beyond those typically considered by individual clinicians. By contrast, individual doctors are confined to their specific field of expertise and can only propose treatment options within their own discipline.

Therefore, the aim of this study is to learn from the MDT decision-making process to develop treatment recommendation method capable of providing optimal, patient-specific therapeutic strategies, thereby supporting clinical decision-making.

Treatment decision-making method based on dual-level embedding representation

This subsection introduces the treatment decision-making method based on dual-level embedding representation. Among them, the word level embedding takes into account the semantic relationships and similarities between words, and the sentence level embedding considers the semantic and contextual information of the sentence⁴⁴. This article adequately represents the intra- and inter-sentence information in the clinical text through both word and sentence levels.

Word-level embedding representation

SMedBERT⁴⁵ was used for word-level embedding representation in this study. The model is an improvement model upon BERT, as it leverages knowledge graphs to inject knowledge into BERT so as to enhance the language model’s understanding capabilities. To be specified, during the pre-training process, incorporating the entity types of entities present in medical texts along with their related neighboring entities deepens the model’s understanding of complex entities and their associated relations in the medical text, thereby enhancing the model’s representation capability.

Given that the author of SMedBERT⁴⁵ have not publicly disclosed a knowledge graph, this research utilized Chinese Medical Knowledge Graph (CMeKG)⁴⁶ to acquire entities, entity types, and relations from collected medical record texts. The framework of SMedBERT based on CMeKG is illustrated in Fig. 2. CMeKG is a comprehensive knowledge graph in the medical domain, encompassing 11,076 diseases, 18,471 drugs, 14,794 symptoms, and 3,546 clinical techniques. With the help of CMeKG, 8,832 triplets, 5,160 entities, 16 kinds of relations and 9 kinds of entity types have been extracted from the collected medical texts. Table 4 demonstrates the extracted relations and entity types. These abundant sources of knowledge will contribute to the comprehensive and accurate representation of words, especially complicated entities.

Fig. 2.

The framework of the SMedBERT based on CMeKG⁴⁵.

Table 4.

Relations and entity types inventory.

Relation	Content
	Clinical manifestations, Adverse reactions, Tests, Complications, Specifications, Therapeutics, Causes, Dosage, Ingredients, Indications, Treatments, Efficacy, Precautions, Associated Diseases, Associated Symptoms, Drug interactions, Functions, Diseases, Symptoms, Drug interactions, and Functions
Entity type	Disease, Body, Symptom, Medical Procedure, Medical Device, Drug, Department, Microbiological, Medical test item

The embedding representation of entities and relations differs from the embedding representation of words obtained through tokenization of text. The words in the text are represented by token embedding, segment embedding, and location embedding. Whereas, the embedding representations of entities and relations are yielded by using the TransR algorithm⁴⁷, which analyzes their semantic correlations within the knowledge graph to derive their embedding representations. The embedding of entity types is obtained from the embeddings of entities. In this study, the average of the embeddings of entity within each entity type is utilized as the embedding representation for entity type. The incorporation of entity type embedding enhances the learning of neighboring entities in subsequent steps to a great extent⁴⁵.

Considering substantial interconnectedness of entities in a knowledge graph, it is crucial to carefully choose relevant neighboring entities for text analysis. This selection process serves to reduce computational burden and minimize the injection of knowledge noise. To accomplish this, we employ PageRank⁴⁸ to compute the weights of neighboring entities for each entity, enabling us to identify the most influential one.

In order to take advantage of the various embeddings mentioned above, SMedBERT incorporates neighbor entity type attention, neighbor entity node attention to learn the structural semantic knowledge. Moreover, the masked neighbor module and masked mention module are added to facilitate the information interaction between mentions and neighbors⁴⁵. Therefore, the SMedBERT can represent medical text with knowledge.

Based on the original configuration of the SMedBERT model parameters, this study performed 50 epochs of fine-tuning on the collected medical record data to generate word vector representations for the medical text. Figure 3 illustrates the change in losses during the fine-tuning process. As can be found in Fig. 3, the loss decreases as epoch increases and finally levels off around 0.10, which indicates that the trained model developed in the study has been able to well represent the medical text. In addition, Figs. 4 and 5 show the number of words and entities in the sample, respectively.

Fig. 3.

Loss during fine-tuning process.

Fig. 4.

Frequency distribution of words in MDT and non-MDT samples (Note: 512 is the max sequence length.).

Fig. 5.

Frequency distribution of entities in MDT and non-MDT samples.

According to Figs. 4 and 5, the following points are found. (1) The number of words in most of the samples in MDT and non-MDT have not exceeded 512. Exceeding 512 means that some information in the word vector representation of the text cannot be represented due to the limitation of the text length. (2) Fig. 5 reveals that a substantial number of entities are accurately identified from the samples using CMEKG, indicating a commendable performance in entity recognition. The integration of specialized medical knowledge enhances the model’s comprehension capabilities. (3) The distribution of words and entities shows a similar pattern between the MDT and non-MDT groups. This provides a foundational basis for recommending treatment patterns for the non-MDT group based on insights gained from the MDT group.

Sentence-level embedding representation

The co-training drove retrieval oriented masking (coROM)⁴⁹, which was initially developed for paragraph retrieval, has been extended to support a wide range of applications, including sentence vector representation, dual-sentence text similarity, and query & multiple candidate documents similarity ranking. coROM was trained under 959,526 passages in medical field of the multi-domain Chinese dataset for passage retrieval (Multi-CPR)⁵⁰. This model demonstrates strong generalization capabilities as a result of being trained on a large-scale training corpus. Furthermore, the model benefits from high-consistency annotations (annotations with more than 80% consistency from at least 5 independent annotators), which ensure its reliability. In general, coROM performs well in capturing semantic relevance between sentences and provides powerful representation of contextual information from the sentence embedding level.

The distribution of the number of sentences contained in the samples of MDT and non-MDT group is shown in Fig. 6. From the figure, it can be observed that the majority of samples have fewer than 25 sentences, with a maximum number of sentences not exceeding 50. In this study, a maximum limit of 50 sentences was set. Compared to the word level embedding representation, the sentence embedding representation ensures that there is no information loss in the text representation.

Fig. 6.

Frequency distribution of sentences in MDT and non-MDT samples.

A splicing method based on an attention mechanism

To achieve a comprehensive representation of text in this study, an attention mechanism⁵¹ is employed to combine word embedding and sentence embedding. The splicing method not only efficiently integrates the two levels of embedding, but also considers the importance of each level of embedding.

Firstly, we employ the attention mechanism to concatenate the dual-level embeddings, addressing the limitation of disregarding the importance of different embedded sentences or words when directly concatenating. Subsequently, a convolutional neural network-static (CNN-static) proposed by Kim⁵² with the capability to handle high-dimensional data processes concatenated text representations to obtain the optimal treatment recommendation. The detailed computations are summarized as follows.

The attention weights under the word and sentence levels and the splicing of dual-level embeddings are calculated with the help of the following Eq. 

1

where and are attention weights vector for words and sentences respectively, and are the embeddings of word level and sentence level respectively, and is the activation function. Words or sentences with large attention weights are important words or sentences in the text and have a significant impact on classification decisions. and are learnable vector parameters. H is the textual representation that integrates word and sentence embeddings. It will be fed into the CNN-static to obtain text classification i.e., treatment recommendation.

Model performance evaluation

We assess model performance from two perspectives, namely prediction performance and association with survival outcomes. The MDT samples are utilized to train the treatment recommendation model. The 5-fold cross-validation is employed to evaluate the prediction performance of the model. Since treatment recommendation is essentially a classification problem, the confusion matrix, accuracy, precision, recall, and F1-score are used to assess prediction performance^4,22.

To assess the association between the treatment by recommendation model and patient survival, the model concordance as an exposure variable²⁷ is introduced to analyze the prognosis in the non-MDT patient group. For each patient in the non-MDT group, we apply the trained model to their EMR data to generate a recommended treatment. We then compare this model-generated recommendation with the actual treatment the patient receives. Patients whose real-world treatment matches the model recommendation are classified as ‘model-concordant’; whereas those whose treatment differs are classified as ‘model-nonconcordant’. Survival outcomes are subsequently compared between these two groups. Considering that only 6 years of EMRs are available, we select 1-year, 3-year survival and 5-year survival as assessment indicators⁴. Furthermore, the Kaplan-Meier method and multivariable Cox proportional hazards regression are employed to construct survival curves and to assess statistical associations between model concordance and survival outcomes.

Three-level explanation

Interpreting models in the medical domain is vital for both facilitating understanding among non-experts and improving the reliability and trustworthiness of the models. This section focuses on providing comprehensive explanations at three levels: word, phrase, and sentence. These explanations are achieved by leveraging attention mechanisms and attention flows⁵³. Important words and sentences are rapidly acquired through attention mechanisms, while crucial phrases are obtained by analyzing the relationships of attentions between different layers by attention flow.

First of all, the words and sentences that are important in the text representation are identified according to the attention value in Eq. (1), the larger the attention value the more important the word or sentence. Then, based on the attention flows proposed by DeRose, et al. ⁵³, the important words are tracked layer by layer. Specifically, the important words for each layer are selected based on multi-head attention from the subsequent layer to the current one. The important words in the last layer have already been obtained in the first step using the attention mechanism. The other layers select the words that have the greatest impact on the important words in the previous layer based on the multi-head attention as the important words in that layer. The relevant calculation formula is shown below.

2

where is the j-th attention head, which used to show the word dependencies between layers. is the embedding vectors of n tokens at layer l, are projection matrices to obtain queries and keys. We only consider words at layer l that have a significant influence ( exceeds the threshold ) on important words at layer l + 1 and count the number of heads that have a significant influence among all attention heads for that word in layer l. The greater the number of attention heads with significant impact, the stronger the role of the l-th layer words in determining the importance of the l + 1-th layer’s important words. This, in turn, increases the likelihood of these words being identified as important.

Finally, the important words from each layer are selected to form phrases based on the attention flows, which provides a phrase-level interpretation of the model results. Thus, the three levels of word, phrase and sentence interpretation are completed. The three levels of interpretation corroborate and complement each other.

Results

The treatment recommendation model was trained on the MDT samples and its prediction performance was evaluated. The performance of the trained treatment recommendation model was assessed using non-MDT samples to examine the association between model concordance and patient survival outcomes. For recommended treatment plan, we leveraged attention mechanism and attention flows to provide explanations at three levels: word, phrase, and sentence. Finally, we conducted ablation experiments to demonstrate the rationality and superiority of the model’s structure.

Performance of the model

The treatment recommendation model proposed in this study underwent a comprehensive evaluation from the perspectives of both prediction performance and associations with survival outcomes. During the evaluation process, we first assessed the model’s predictive capabilities, and measured its accuracy and reliability in predicting suitable treatment options based on the input data. Subsequently, we focused on evaluating the association between model concordance and patient survival.

Prediction performance

Based on word embedding and sentence embedding representations by methods in Sect. 3.2.1 and 3.2.2, CNN-static with an attention mechanism was used to predict the treatment plan. The CNN-static⁵² was equipped with 200 filters of sizes [10, 15, 20], with dropout rate of 0.4, a learning rate set to 0.01, and it was trained for 500 epochs. To assess the model’s performance, we conducted 5-fold cross-validation. Figure 7 shows the results of the normalized confusion matrix obtained during cross-validation, which allows us to fairly compare the model’s performance across various categories despite differences in sample sizes. Table 5 summarizes the prediction results for each fold. To further account for variability across categories, Table 6 reports the performance metrics for each category along with 95% confidence interval (CI).

Fig. 7.

Results in 5-fold cross validation (Note: All values in the figure are represented as percentages. ‘Statistics’ represents the aggregated confusion matrix obtained from 5-fold cross-validation. The meaning of the labels: 0 refer to ‘Surgery’, 1 refer to ‘Surgery after neoadjuvant therapy’, 2 refer to ‘Chemotherapy’, 3 refer to ‘Radiotherapy’, 4 refer to ‘Chemoradiotherapy (Sequential/Concurrent)’, 5 refer to ‘Chemotherapy + immunotherapy/targeted therapy’.).

Table 5.

Prediction performance of the proposed model through 5-fold cross-validation.

Fold 1	Accuracy	Precision	Recall	F1-score
	0.8783	0.8479	0.8720	0.8554
Fold 2	0.8609	0.8286	0.8600	0.8391
Fold 3	0.8889	0.8422	0.8797	0.8567
Fold 4	0.9217	0.8955	0.9102	0.9007
Fold 5	0.9123	0.8798	0.8976	0.8873
Overall	0.8924 (0.8616–0.9232)	0.8588 (0.8243–0.8934)	0.8839 (0.8590–0.9088)	0.8679 (0.8365–0.8993)

Note: In each fold, the evaluation metrics are computed using macro-average, and “Overall” represents the mean value of each metric across all folds, along with its 95% CI.

Table 6.

Prediction performance of the proposed model across treatment categories (5-fold cross-validation).

Treatment category	Precision	Recall	F1	Accuracy
SUR	0.9230 (0.8686–0.9774)	0.8857 (0.8333–0.9381)	0.9032 (0.8632–0.9433)	0.9549 (0.9342–0.9755)
SURANT	0.9022 (0.8332–0.9712)	0.8636 (0.8347–0.8925)	0.8819 (0.8429–0.9208)	0.9497 (0.9242–0.9753)
CHE	0.7011 (0.5735–0.8287)	0.8455 (0.7791–0.9118)	0.7635 (0.6721–0.8550)	0.9601 (0.9507–0.9695)
RAD	0.7262 (0.6400–0.8124)	0.8511 (0.7989–0.9033)	0.7820 (0.7282–0.8358)	0.9705 (0.9581–0.9828)
CHE (S/C)	0.9235 (0.8834–0.9636)	0.9541 (0.9405–0.9677)	0.9383 (0.9159–0.9607)	0.9757 (0.9667–0.9847)
CHE + IMM/TAR	0.9769 (0.9350–1.0000)	0.9034 (0.8444–0.9624)	0.9383 (0.8943–0.9822)	0.9739 (0.9498–0.9981)

From the results presented in Fig. 7, Tables 5 and 6, several conclusions can be drawn in the following. (1) The model achieves a high recall of 0.9541 for the class ‘Chemoradiotherapy (Sequential/Concurrent)’ and 0.9034 for ' Chemotherapy + immunotherapy/targeted therapy ‘, indicating that it successfully identifies most patients who should receive these treatments. Moreover, the precision for ‘Chemotherapy + immunotherapy/targeted therapy’ (0.9769) and ‘Chemoradiotherapy (Sequential/Concurrent)’ (0.9235) suggests that the model’s recommendations in these categories are highly accurate. The CIs in Table 6 further confirm the reliability and robustness of these results. (2) The model shows relatively lower precision in ‘Chemotherapy’ (0.7011) and ‘Radiotherapy’ (0.7262), though their recall values remain high (0.8455 and 0.8511, respectively). This indicates that while the model can identify most patients needing these treatments, it sometimes produces false positives. This performance may be attributed to the smaller sample sizes in these categories, limiting the model’s learning. Nevertheless, Table 6 shows that the model still maintains high accuracy for these categories (0.9601 for Chemotherapy and 0.9705 for Radiotherapy), demonstrating its potential for practical use. (3) Overall, across the five folds, the model achieves macro-average accuracies ranging from 0.8609 to 0.9217 (Table 5), with an overall indicator of 0.8924 ± 0.0222. Similarly, macro-average precision, recall, and F1-score are all above 0.85, with modest variance between folds. Table 6 further confirms that the model’s accuracy remains consistently high across all treatment categories, with narrow 95% CI, demonstrating strong generalization ability and robustness.

Survival analysis

We stratified non-MDT patients into model-concordant and model-nonconcordant groups to compare their survival outcomes. To determine model concordance, we applied the trained model to each patient’s EMR data to generate a recommended treatment and compared it with the treatment actually received. Patients whose treatments matched the model recommendation were classified as ‘model-concordant’, while those with differing treatments were classified as ‘model-nonconcordant’. This evaluation approach, which considers model concordance as the exposure factor, has been widely utilized for outcome assessments²⁷. Time zero was defined as the date of diagnosis, and patients were followed until the occurrence of the event (death) or censoring.

Among the non-MDT patient population, 23.37% of the patients adopted treatment plans concordant to the model recommendations. Table 7 summarizes the baseline characteristics of non-MDT patients stratified by model concordance. The baseline characteristics were comparable between the model-concordant and model-nonconcordant groups in terms of age, sex, disease stage, and disease origin. Table 8 presents the one-year, three-year, and five-year survival outcomes of the model-concordant group and the model-nonconcordant group, and the chi-square test was conducted to analyse the differences in survival between the two groups. The results indicate that the model-concordant group exhibits higher observed survival rates, with statistically significant differences at a 99% confidence level in the one-year, three-year, and five-year survival periods. These findings highlight a statistically significant association between model concordance and patient survival in the studied dataset.

Table 7.

Baseline characteristics of non-MDT patients stratified by model concordance.

Characteristic	Model-concordant (n = 455)	Model-nonconcordant (n = 1492)	p-value
Age, years (mean ± standard deviation)	63.9 ± 9.4	64.3 ± 10.1	0.4353
Sex, n (%)			0.7326
Male	284 (62.4%)	918 (61.5%)
Female	171 (37.6%)	574 (38.5%)
Disease stage, n (%)			0.7266
Stage IIIA	183 (40.2%)	569 (38.1%)
Stage IIIB	147 (32.3%)	498 (33.4%)
Stage IIIC	125 (27.5%)	425 (28.5%)
Disease origin, n (%)			0.4558
De novo	394 (86.6%)	1271 (85.2%)
Recurrent/progression	61 (13.4%)	221 (14.8%)

Note: Continuous variables are presented as mean ± standard deviation and compared using the Welch t-test. Categorical variables are presented as number (percentage) and compared using the chi-square test.

Table 8.

One-year, three-year, and five-year survival outcomes of the model-concordant group and the model-nonconcordant group.

The model-concordant group	One-year survival	Three-year survival	Five-year survival
	49.01%	23.96%	11.43%
The model-nonconcordant group	32.51%	8.71%	2.82%
(p-value)	24.87(6.12e-07)	73.82(8.55e-18)	54.44(1.61e-13)

The Kaplan-Meier survival analysis³³ revealed a highly significant difference between the model-concordant and model-nonconcordant groups, with a log-rank p-value of 1.2804 × 10⁻¹⁴ shown in Fig. 8. The estimated one-year, three-year, and five-year survival probabilities for the model-concordant group were 46.4% (95% CI, 41.6–50.8%), 24.0% (95% CI, 20.0–27.8%), and 11.4% (95% CI, 8.5–14.3%), respectively, whereas for the model-nonconcordant group, the corresponding survival probabilities were markedly lower at 33.2% (95% CI, 30.8–35.6%), 8.7% (95% CI, 7.3–10.1%), and 2.8% (95% CI, 2.0–3.7%). These trends are consistent with the data presented in Table 8 and indicate that patients in the model-concordant group experienced substantially better survival outcomes over time. The Kaplan-Meier curves further illustrate that the divergence in survival probabilities between the two groups becomes most pronounced between 500 and 2000 days, and the number of patients at risk provided at each time point reinforces the reliability of these estimates.

Fig. 8.

Kaplan-Meier Survival Curve: A comparison between model-concordant group and model-nonconcordant group.

Multivariable Cox proportional hazards regression analysis was performed to assess whether the association between model concordance and survival remained independent of other clinical factors, including disease origin, stage III subtype, age, and sex. The results shown in Table 9 demonstrated that model concordance was a strong independent predictor of survival, with a hazard ratio (HR) of 0.67 (95% CI, 0.60–0.76, p < 0.001), indicating that patients in the concordant group had a substantially lower risk of adverse outcomes. In addition, several other covariates also exhibited statistically significant associations with survival. Specifically, patients with recurrent or progressive disease had a higher risk of adverse outcomes compared to those with de novo stage III disease (HR 1.12, 95% CI 1.01–1.24, p = 0.03), and higher stage III subtypes were linked to progressively higher risk (stage IIIB vs. IIIA: HR 1.18, 95% CI 1.05–1.33, p = 0.007; stage IIIC vs. IIIA: HR 1.42, 95% CI 1.25–1.61, p < 0.001). Age was also a modest but significant risk factor (HR 1.02 per year, 95% CI 1.01–1.03, p < 0.001), whereas sex did not reach statistical significance (HR 1.08, 95% CI 0.97–1.20, p = 0.15). These results suggest that while model concordance remains the strongest independent predictor, survival outcomes are also influenced by disease characteristics and patient age.

Table 9.

Multivariable Cox proportional hazards analysis of factors associated with survival in non-MDT patients.

Variable	HR	95% CI	p-value
Model concordant	0.67	0.60–0.76	< 0.001
Disease origin (recurrent/progression vs. de novo)	1.12	1.01–1.24	0.03
Stage IIIB vs. IIIA	1.18	1.05–1.33	0.007
Stage IIIC vs. IIIA	1.42	1.25–1.61	< 0.001
Age (per year)	1.02	1.01–1.03	< 0.001
Sex (male vs. female)	1.08	0.97–1.20	0.15

Together, the Kaplan–Meier analysis and multivariable Cox regression provide complementary evidence that model concordance is associated with markedly improved long-term survival. This underscores the potential utility of the model not only in predicting outcomes but also in stratifying patients for prognostic assessment, highlighting its clinical relevance in guiding management strategies.

Model interpretation

This study elucidates the model results at the word, phrase, and sentence levels, aiming to explore and reveal the rationale for treatment recommendations. During the interpretation process, we randomly selected two samples under different treatment plans to present the model interpretation results. Following the approach described in Sect. 3.4, the obtained key words, phrases, and sentences are demonstrated in Figs. 9–10.

Figs. 9.

The explanation for the treatment of chemotherapy + immunotherapy/targeted therapy.

Fig. 10.

The explanation for the treatment of surgery after neoadjuvant therapy.

Figs. 9–10 provide a comprehensive insight into the model’s interpretability. These figures first present the important words and sentences based on the attention values. Subsequently, it highlights the key words from the last two layers using attention flows (although the study considered key words from the last five layers, only the last two are displayed due to space limitations). Finally, the model leverages these key phrases and important sentences to elucidate treatment options, such as the potential for surgery, the need for adjuvant therapy, and the suitability for immunotherapy. To sum up, with the help of Figs. 9–10, the basis of the model’s recommendation results can be clearly understood. These key sentences and phrases serve as a reasonable basis for decision-makers to understand the model’s results, which provides valuable insights into the patient’s condition and supports the treatment recommended by the model.

Ablation study

To demonstrate the superior performance of the model and showcase the roles of each module in prediction, we conducted ablation experiments. That is to say, we examined the advantages of dual-level embedding, the strengths of knowledge-enhanced word embedding, and the benefits of splicing word embedding and sentence embedding through attention mechanisms. The observed performance improvements are based on the current dataset.

Effect of dual-level embedding

In the embedding representation of clinical texts, we employed a dual-level embedding approach. To investigate the impact of dual-level embeddings on recommendation performance, we compared the model’s effectiveness under two scenarios: utilizing only word-level embedding (SMedBERT) and employing only sentence-level embedding. The comparative results are presented in Table 10. The table not only provides the model’s performance across different categories but also showcases the overall model performance under macro-average metrics. Comparing Table 10 with Table 6 reveals that the model’s prediction performance is poorer when predicting categories such as chemotherapy and radiotherapy under either sentence or word embeddings alone. This is primarily attributed to insufficiently representing information in the text, leading to suboptimal performance when predicting these easily confused categories. By comparing Table 10 with Table 5, it is evident that under macro-average evaluation metrics, the dual-level embedding approach yields a performance improvement of at least 5% compared to using a single embedding representation method. This further validates that dual-level embeddings can comprehensively represent textual information, demonstrating significant advantages over single-level embedding.

Table 10.

Effect of dual-level embedding.

Methods	Category	Accuracy	Precision	Recall	F1-score
Only word embedding by SMedBERT	SUR	0.9460 (0.9254–0.9665)	0.9199 (0.8634–0.9763)	0.8480 (0.7995–0.8965)	0.8818 (0.8417–0.9219)
	SURANT	0.9373 (0.9107–0.9639)	0.8750 (0.8156–0.9344)	0.8279 (0.8073–0.8485)	0.8503 (0.8183–0.8824)
	CHE	0.9442 (0.9387–0.9496)	0.6085 (0.5145–0.7024)	0.7459 (0.6544–0.8374)	0.6694 (0.5797–0.7591)
	RAD	0.9477 (0.9346–0.9607)	0.5553 (0.4491–0.6615)	0.7022 (0.5979–0.8066)	0.6189 (0.5184–0.7194)
	CHE (S/C)	0.9512 (0.9418–0.9605)	0.8417 (0.7971–0.8863)	0.9144 (0.8732–0.9557)	0.8764 (0.8371–0.9156)
	CHE + IMM/TAR	0.9634 (0.9345–0.9923)	0.9691 (0.9087–1.0000)	0.8610 (0.8016–0.9203)	0.9113 (0.8599–0.9627)
	Overall	0.8448 (0.8097–0.8800)	0.7949 (0.7455–0.8443)	0.8166 (0.7769–0.8562)	0.8013 (0.7562–0.8465)
Only sentence embedding	SUR	0.9457 (0.9250–0.9663)	0.9191 (0.8631–0.9750)	0.8458 (0.7954–0.8962)	0.8802 (0.8401–0.9203)
	SURANT	0.9335 (0.9028–0.9642)	0.8668 (0.8027–0.9309)	0.8212 (0.7827–0.8597)	0.8428 (0.8018–0.8838)
	CHE	0.9421 (0.9367–0.9476)	0.5974 (0.4773–0.7174)	0.7459 (0.6544–0.8374)	0.6618 (0.5551–0.7684)
	RAD	0.9544 (0.9400–0.9687)	0.6176 (0.5606–0.6746)	0.7022 (0.5979–0.8066)	0.6552 (0.5928–0.7176)
	CHE (S/C)	0.9491 (0.9398–0.9584)	0.8409 (0.7975–0.8843)	0.9078 (0.8813–0.9342)	0.8729 (0.8406–0.9052)
	CHE + IMM/TAR	0.9562 (0.9259–0.9866)	0.9320 (0.8786–0.9854)	0.8600 (0.8013–0.9187)	0.8944 (0.8404–0.9485)
	Overall	0.8405 (0.8036–0.8773)	0.7956 (0.7492–0.8420)	0.8138 (0.7726–0.8551)	0.8012 (0.7567–0.8457)

Effect of splicing through attention mechanisms

In the context of dual-level embeddings, we employ an attention mechanism to merge the two levels of embeddings, rather than directly concatenating them. This approach serves two purposes: on the one hand, it facilitates the interpretation of recommendation results; on the other hand, it enhances the efficiency of information integration. To investigate the impact of attention-based integration compared to direct concatenation, we compared their performance in the model. Table 11 provides a detailed presentation of the model’s prediction performance across different categories and its overall performance under macro-average metrics when using direct splicing. Comparing this with Table 6 reveals a significant improvement in the model’s prediction performance for each category upon introducing the attention mechanism. Further comparison with Table 5 indicates an overall performance improvement of approximately 1% due to the introduction of the attention mechanism. These results validate the superiority of the attention mechanism in more effectively integrating information from two hierarchical levels.

Table 11.

Effect of splicing through attention mechanisms.

Methods	Category	Accuracy	Precision	Recall	F1-score
Direct splicing	SUR	0.9548 (0.9341–0.9754)	0.9230 (0.8686–0.9774)	0.8857 (0.8333–0.9381)	0.9032 (0.8632–0.9433)
	SURANT	0.9478 (0.9184–0.9772)	0.9002 (0.8295–0.9710)	0.8557 (0.8108–0.9006)	0.8768 (0.8294–0.9241)
	CHE	0.9582 (0.9444–0.9720)	0.6823 (0.5468–0.8179)	0.8396 (0.7607–0.9185)	0.7509 (0.6418–0.8600)
	RAD	0.9704 (0.9580–0.9827)	0.7262 (0.6400–0.8124)	0.8511 (0.7989–0.9033)	0.7820 (0.7282–0.8358)
	CHE (S/C)	0.9704 (0.9580–0.9827)	0.9216 (0.8799–0.9634)	0.9302 (0.9087–0.9517)	0.9257 (0.8999–0.9514)
	CHE + IMM/TAR	0.9686 (0.9416–0.9957)	0.9497 (0.9013–0.9982)	0.9034 (0.8444–0.9624)	0.9258 (0.8750–0.9766)
	Overall	0.8851 (0.8494–0.9208)	0.8505 (0.8107–0.8904)	0.8776 (0.8448–0.9105)	0.8607 (0.8231–0.8983)

Effect of knowledge infusion

In the word-level embedding representation, a knowledge graph (CMeKG) is utilized to enhance the capabilities of the BERT model. To study the impact of knowledge infusion on model performance, we compared the results with and without the incorporation of the knowledge graph. Table 12 presents the model’s prediction performance across various categories and its overall performance under macro-average performance in the absence of a knowledge graph. Compared to Table 5, it can be observed that knowledge injection results in approximately a 1% performance improvement. Although this improvement may seem modest, contrasting it with Table 6 reveals a substantial enhancement in the model’s prediction performance for chemotherapy and radiotherapy. This validates that knowledge infusion indeed enhances the performance of the word representation method, and provides more accurate information for the model.

Table 12.

Effect of knowledge infusion.

Method	Category	Accuracy	Precision	Recall	F1-score
Word embedding by BERT without knowledge graph	SUR	0.9547 (0.9315–0.9779)	0.9241 (0.8731–0.9751)	0.8897 (0.8166–0.9628)	0.9050 (0.8669–0.9431)
	SURANT	0.9477 (0.9212–0.9742)	0.8886 (0.8222–0.9550)	0.8692 (0.8391–0.8993)	0.8782 (0.8387–0.9176)
	CHE	0.9634 (0.9546–0.9722)	0.7256 (0.6156–0.8356)	0.8436 (0.7806–0.9067)	0.7777 (0.7045–0.8508)
	RAD	0.9651 (0.9519–0.9783)	0.6743 (0.6027–0.7458)	0.8511 (0.7989–0.9033)	0.7512 (0.6981–0.8044)
	CHE (S/C)	0.9651 (0.9574–0.9728)	0.9194 (0.9028–0.9361)	0.9017 (0.8821–0.9214)	0.9104 (0.9020–0.9187)
	CHE + IMM/TAR	0.9703 (0.9451–0.9954)	0.9611 (0.9141–1.0000)	0.9069 (0.8545–0.9592)	0.9329 (0.8881–0.9777)
	Overall	0.8831 (0.8573–0.9090)	0.8489 (0.8145–0.8832)	0.8770 (0.8517–0.9024)	0.8592 (0.8281–0.8904)

In summary, through the adoption of dual-level embedding and knowledge injection in the model’s architecture, along with the incorporation of attention mechanisms, we successfully develop a rational and effective model that can efficiently process information and achieve remarkable performance improvements in the treatment recommendation task.

Conclusions

Lung cancer, as the leading cause of death in the world, brings a heavy economic burden and serious physical damage to people. MDT can provide the best treatment plan for cancer patients and enhance their prognosis. However, MDT consumes too much time and energy of specialist doctors in treatment decisions, which is not applicable in countries with limited medical resources. This paper proposes develops an interpretable MDT treatment recommendation approach to provide optimal treatment options for patients with stage III NSCLC with high heterogeneity. A dual-level embedding approach that integrates word embedding and sentence embedding for clinical text representation is proposed in this study. And a comparative analysis with word embedding-only and sentence embedding-only further illustrates that the proposed approach considers both local and global information and more adequately represents clinical text. Subsequently, taking into account the heterogeneity among patients, the basis for treatment decision making is dissected in depth at three levels: word, phrase and sentence, identifying key factors influencing treatment. Specially, attention flow is employed to consider the relationship between attentions across multiple layers, and important words and sentences are screened out through attention mechanisms. Finally, using stage III NSCLC in China as an example, it is concluded not only that the proposed model can highly simulate MDT treatment recommendations, but also that the recommended treatments show potential to be associated with improved survival prognosis.

Additionally, there are also some limitations. First, the evaluation of the proposed model was restricted to treatment decision-making for stage III NSCLC, and only textual information extracted from EMRs was included. Second, as a retrospective study, potential unmeasured confounding factors may have influenced the observed results, which could not be fully controlled. Third, the dataset was derived from only two hospitals, which may limit the external validity and generalizability of the findings. Finally, imbalances in the distribution of treatment categories existed in some cases, which might have affected the robustness of the model’s performance.

Future studies will aim to address these limitations. First, expanding the model evaluation to additional cancer types and incorporating structured and unstructured EMR data will broaden the applicability and improve decision support. Second, prospective study designs and advanced statistical methods for confounding control (e.g., propensity score matching or causal inference techniques) should be employed to mitigate the influence of unmeasured confounders. Third, validation on multi-center, large-scale, and geographically diverse datasets is necessary to enhance the generalizability and external validity of the model. Finally, data balancing techniques and more representative sampling strategies will be explored to reduce treatment category imbalances and further improve model stability and fairness.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Ziyu Chen: Conceptualization, Methodology, Writing-Original Draft. Naijie Chai: Visualization, Investigation, Writing-Review & Editing. Jianqiang Wang: Writing, Reviewing, Editing-Supervision. Xiaokang Wang: Data Curation, Software, Visualization-Review & Editing. Zhenqian Fan: Project administration, Writing-Review & Editing.

Funding

The authors thank the editors and anonymous reviewers for their helpful comments and suggestions. This work was supported by the Natural Science Foundation of Shaanxi Province (Nos. 2025JC-YBQN-996 and 2025JC-YBQN-988), Research Project of Humanities and Social Sciences of the Ministry of Education (Nos. 25YJC630007 and 25YJCZH011), the Postdoctoral Fellowship Program of CPSF under Grant Number GZC20251119.

Data availability

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.