In the clinical and biomedical fields, electronic health records (EHRs) have become valuable resources in recent years [1]. In order to protect the privacy of patients whose data are used for secondary purposes, regulations or laws have been established requiring the removal of protected health information (PHI) from records before they can be disseminated. In settings, such as research, obtaining explicit consent to access personal data may be impractical or impossible. As a result, the deidentification step becomes a critical data processing step to protect the privacy of individuals. However, the study by Cannon and Lucci [2] indicated that up to 65% of important clinical information is recorded in unstructured texts in medical reports written by medical personnel. Compared with structured data, which can be deidentified by encrypting private patient information fields, the deidentification of unstructured data is more challenging.

Because manual deidentification of large volumes of EHRs is time-consuming and error-prone, automated methods are needed for large-scale deidentification of unstructured clinical data. Compared with Asian countries, such as Japan and China, where the majority of medical records are written monolingually in their native languages, medical records in Taiwan are frequently written in a mixture of Chinese and English. The example sentence “前夫mk1300309 married > mk136從婆家搬出來 > mk139離婚” (got married to my ex-husband on 2141/03/09 > moved out from the ex-husband’s family’s home in 2147 > divorced in 2150) demonstrates how a physician wrote in English and suddenly switched to Chinese. The sentence also includes a transliteration (“mk”) of “民國,” which refers to the Republic of China calendar. This phenomenon is referred to as code alternation, which occurs when a bilingual speaker uses more than one language in a single utterance [3]. Code switching refers to code alternation at or above clause level, while code mixing refers to code alteration below clause level. Some examples of code-switched and code-mixed narratives in discharge summaries are shown in Multimedia Appendix 1. Throughout this work, we will use code mixing to refer to the above phenomenon.

In addition to our specific context, it is important to note that code mixing in clinical records is observed in various multilingual health care settings globally. For instance, Alqurashi [4] conducted research on the use of code-mixed language in private hospitals in Saudi Arabia, highlighting the prevalence of this practice in health care communication. Furthermore, Dowlagar and Mamidi [5] demonstrated that the use of code-mixed utterances, where native languages are blended with English, is a common and natural occurrence among doctors and patients. Beyond this, code mixing in clinical notes extends beyond English-dominant countries. For example, Keresztes [6] investigated how Hungarian physicians are influenced by the English language in their writing of cardiology discharge reports, while Karuthan [7] showed that nurses in Malaysia tend to use “Manglish,” a mix of Malayalam and English, when writing nursing documents.

This global prevalence of code mixing highlights the importance of addressing narratives written in this manner. Unfortunately, most existing clinical natural language processing (NLP) techniques are primarily designed for monolingual texts, mainly due to limited research resources. While our previous work [8] has highlighted the potential of pretrained language models (PLMs) in mitigating issues related to code mixing during deidentification, the precise mechanism at play remains incompletely understood. To bridge this knowledge gap, there is a pressing need for a code-mixed deidentification data set that can comprehensively evaluate the performance of state-of-the-art NLP models and unravel their underlying mechanisms.

Previous studies [9,10] have demonstrated that state-of-the-art neural network models can effectively exploit strong name regularity and high mention coverage to achieve superior performance. Name regularity refers to the consistency and predictability of the structure and pronunciation of a named entity type within a particular language. For example, person names typically adhere to the “FirstName LastName” format. A high mention coverage denotes that a large proportion of mentions in the test set have been previously observed in the training set.

Our study aims to explore the impact of these properties on the performance of PLMs in recognizing PHI within code-mixed sentences. We hypothesize that the decision-making processes of the models are influenced by these factors, even when dealing with code-mixed text. However, they may underperform when faced with weak regularity, such as unseen mentions or out-of-vocabulary (OOV) words. Additionally, they may solely memorize popular mentions instead of learning generalization knowledge, which could limit their generality and performance. To investigate the decision-making process of these models and gain insights into their ability to recognize code-mixed PHI, we curated a novel code-mixed deidentification data set, trained and evaluated the performance of state-of-the-art PLMs on the unique data set, and employed methods to interpret their results to validate our hypothesis and gain a deeper understanding of how these models recognize code-mixed PHI.

In summary, this study has the following 3 major contributions:

We obtained discharge summaries sampled from the psychiatry section of a medical center, which were used to extract depressive symptoms with text mining approaches [39,40]. From these data, we compiled a code-mixed deidentification corpus of 1700 discharge summaries. Four annotators were enlisted to annotate the entire corpus, and they followed the same annotation procedure as previously described in our work [8]. To ensure the quality of the annotation, the annotation process began with the annotators individually annotating an identical set of 200 randomly sampled records, using the annotation guidelines provided. Subsequently, a meeting was organized to facilitate discussions among the annotators, addressing any issues or concerns that arose during the annotation process. This iterative process continued until the annotators achieved a strong level of agreement [41]. The interannotation agreement, as measured by κ, was found to be 0.85 after the above procedure. Once this agreement threshold was reached, the remaining unlabeled EHRs were evenly distributed among the 4 annotators for labeling. The training set comprised 1500 discharge summaries along with 297,621 sentences. The test set comprised 200 discharge summaries along with 60,632 sentences. Finally, the principle-based resynthesis method proposed in our previous work [38] was used to generate surrogates, and the entire corpus was rechecked by one of the senior annotators (PTC) to ensure a high level of data consistency and correctness.

The study has been approved by the ethics committee for medical research of the National Institutes of Health (number: EC1090212-E).

After preprocessing the collected data set, we approached the deidentification task by treating it as a named entity recognition problem in which the target entity types are PHI types defined in Table 1. To establish a baseline, we used a dictionary-based method [42,43] and developed 4 BERT-based models for comparison. The dictionary-based approach (DBA) relies on predefined dictionaries and a lookup method. In our implementation, we compiled a dictionary consisting of tokens for all PHI types collected from the training set. The most frequent tag associated with each token estimated on the training set was set as the token’s tag. For the BERT-based models, the recognition task was formulated as a sequential tagging problem by using the BILOU tagging schema. This schema categorizes the tokens as either the beginning (B), within (I), or last (L) of multi-word PHI, as well as identifies non-PHI (O) and single-word PHI (U) tokens. For more in-depth information about the preprocessing steps and the development of the DBA, please refer to Multimedia Appendix 2.

Given the code-mixing characteristics of our corpus, we categorized sentences in our data set based on their code-mixing characteristics into 4 groups: English only, Chinese only, a mixture of Chinese and English, and numeric/symbolic characters. The statistical information for PHI across the 4 groups is presented in Table 2. Overall, it is apparent that the distributions of different PHI types are significantly imbalanced. Without considering numeric/symbolic sentences, the most frequent type of sentences was English only, followed by code-mixed and Chinese only. The “Date” type was most frequently observed in numeric/symbolic sentences, while the “Age” type was most commonly noted in English-only sentences. The “Name” and “Profession” types were most frequently observed in code-mixed sentences, and the “Location” type was evenly distributed across English-only, Chinese-only, and code-mixed sentences. Finally, the “ID” type was most commonly found in Chinese-only sentences.

The AvgCMI_English, AvgCMI_Chinese, and CMIC estimated for the training set, test set, and entire corpus are presented in Multimedia Appendix 3, which highlight the writing convention in Taiwan’s medical records, with the majority of records being written in English. Another interesting writing convention observed was that code mixing was more frequent in Chinese-dominated sentences than in English-dominated sentences. The significant difference between AvgCMI_English and AvgCMI_Chinese highlighted that the frequency of Chinese tokens occurring in English-dominated sentences is much less than the frequency of English tokens occurring in Chinese-dominated sentences. The CMIC of our corpus was 22.21, which is significantly higher than the CMIC of English-Bangla (approximately 5.15) and Dutch-Turkish (approximately 4.13) data sets [45].

The performance comparison of the developed deidentification methods is shown in Table 3, where the DBA indicates the dictionary-based baseline. The remaining configurations include monolingual BERT models (EN-BERT and CH-BERT) and multilingual BERT models (M-BERT and TM-BERT).

The results showed a significantly lower performance of the DBA, highlighting that the dictionary collected from the training set is insufficient to provide reliable deidentification results. To be specific, the micro- and macro-F-scores of the DBA were 0.13 and 0.16, respectively, while the micro- and macro-F-scores of BERT-based methods ranged from 0.90 to 0.94 and 0.63 to 0.79, respectively, which were significantly higher than those of the DBA. Among all BERT-based approaches, M-BERT outperformed the others in terms of both micro- and macro-F-scores. CH-BERT demonstrated a performance comparable to that of M-BERT on various PHI types such as “Name” and “Location.” The micro- and macro-F-scores of TM-BERT were higher than those of EN-BERT but lower than those of CH-BERT and M-BERT, suggesting that translating code-mixed sentences into monolingual text does not provide advantages. TM-BERT’s lower performance was mainly due to translation errors that can cause syntactic or semantic issues during fine-tuning. Some words like abbreviations (“NKUST”) and Chinese names that are PHI mentions cannot be accurately translated. Even context words could alter their meaning, for example, the text “occupation: 英語補習班老師” (Profession: English tutoring teacher) was translated as “佔領: 英語補習班老師” (Domination: English tutoring teacher).

To gain a deeper understanding of the ability of PLMs to recognize different types of PHI mentioned in sentences with different levels of language mixing, we categorized the sentences in our corpus into 3 distinct sentence categories: English-only, Chinese-only, and Chinese-English mixed. We then evaluated the performance of the developed methods for each category. The results are presented in Table 4, focusing on the comparison between EN-BERT, CH-BERT, and M-BERT only. We observed that the distribution of different PHI types varied across the 3 sentence categories, suggesting that the performance of the developed models may be influenced by the prevalence of certain PHI types in their respective sentence categories. Specifically, we found that numeric PHI types, such as “Date” and “Age,” were more prevalent in English-only sentences, while nonnumeric PHI types, including “Name” and “Profession,” were more frequent in code-mixed sentences. This finding explains why EN-BERT performed similarly to multilingual and translation models on numeric PHI types but much worse on nonnumeric types. Additionally, the generally high performance for numeric PHI types indicates that mixing Chinese and English in the same sentence has little effect on their recognitions. Detailed results can be found in Multimedia Appendix 3.

In contrast, the “ID” PHI type occurred most frequently in the Chinese-only sentences, but all BERT-based models performed well in recognizing them. The “Location” PHI type was equally present in all 3 categories, but certain fine-grained types, such as “Region” and “Nationality,” were more challenging due to their limited training instances. Overall, the results showed that M-BERT outperformed the other models in the 3 sentence categories based on the metric of micro-F-score, while CH-BERT had the best performance in terms of macro-F-score in Chinese-only and code-mixed sentences. EN-BERT exhibited comparable performance with M-BERT in English-only sentences, but its performance was inferior to that of CH-BERT and M-BERT in Chinese-only and code-mixed sentences, indicating that these types of sentences are more challenging for EN-BERT. In contrast, both CH-BERT and M-BERT exhibited less effect of the code-mixing issue in their recognition of PHI in code-mixed sentences.

In this subsection, we report sentence error rates by dividing sentences in our corpus into 6 groups according to their CMI levels, as illustrated in Figure 3, to gain insights into the impact of the code-mixing level on PHI recognition. It is evident from the figure that EN-BERT has the highest error rate for all CMI ranges. This aligns with our earlier finding in the previous section demonstrating that EN-BERT struggles with recognizing PHI in code-mixed sentences. Although CH-BERT and M-BERT exhibited lower sentence error rates than EN-BERT, the error rates were still considerably higher for code-mixed sentences (CMI >0), indicating that the recognition of PHI in code-mixed sentences remains a challenging task for the 2 PLMs.

Further evidence supporting this conclusion can be found in Table 5, where we fine-tuned M-BERT on training instances from 1 sentence category and evaluated it on the test set categorized into the 3 categories. From the results, we observed that fine-tuning M-BERT with training instances from a specific sentence category can significantly improve its performance on that category. However, fine-tuning on the code-mixed training instances yielded the best results, which further enabled the model to effectively recognize PHI in both Chinese and English sentences. This finding emphasizes the significance of incorporating code-mixed training instances into the model’s training data.

To assess the effectiveness of the proposed ChatGPT-based framework for recognizing PHI in our code-mixed corpus, we collected a subset of the original data that included sentences containing one or more PHI sets that were not recognized by any of the BERT-based models we developed. This subset was chosen to represent the most challenging cases of the compiled corpus while minimizing the risk of sensitive information exposure. Figure 4 shows the results, with M-BERT used for comparison.

Compared with M-BERT, the proposed ChatGPT-based framework achieved slightly better recall, but lower precision, leading to a lower micro-F-score in the zero-shot setting. Nonetheless, our framework presents an appealing approach to use LLMs for the deidentification task over standard-sized models like BERT. LLMs are notably easier to use and can be controlled through natural language prompts, requiring little to no machine learning expertise. For example, we found only 1 error case for the “Doctor” PHI that both BERT-based models and ChatGPT failed to recognize: “follow-up: (ydh/ptz).” However, we can enhance the prompt shown in Figure 2 by including a statement following the “DOCTOR NAME” definition that lists the abbreviation names of physicians in a hospital, directing ChatGPT to recognize “ptz” as a doctor name: “Here is a list of abbreviated physician names of our hospital. Remember and recognize them from the given text: 1. hjd, 2. ptz, 3. ctc.” This exemplifies the adaptability and flexibility of LLMs for deidentification tasks, making them particularly well-suited for real hospital environments. While the potential for using ChatGPT in deidentification is notable, the model is currently only accessible through an online application programming interface. This renders it unsuitable for use in a hospital setting, where patient data cannot be stored or transmitted to unauthorized external parties.

On the other hand, it crucially highlights the significance of carefully crafted prompts when working with ChatGPT. The constraint statements presented in Figure 1 play a pivotal role in guiding ChatGPT’s responses. When these constraints were removed, we observed that ChatGPT can generate responses that may be unpredictable and go beyond the desired control. Take the text “he was brought to Dr. 莊凱傑’s opd on 3/13 and haldol + anxicam 1 amp was given” as an example. Here, “莊凱傑” represents a Chinese name, which can be translated to “KAI-JIE ZHUANG.” Using the developed prompts without adding the constraint part, both the original and recently released ChatGPT 3.5 models (version from September 27, 2023) could potentially return responses like “MEDICATION: haldol, anxicam” and “DOSE: 1 amp.”

Another issue to consider is the translation of recognized PHI into a language other than that mentioned in the given text. The issue is more prominent in Chinese-majority code-mixed sentences, where the recognized Chinese PHI might be translated into the English counterparts. For example, in the sentence “education: 高中讀中和高中, 僅唸數個月就因跟不上而休學, 大學念正修夜間部商學” (education: high school-Zhonghe senior high school, dropped out after only a few months; College-Business in the night school at Cheng Shiu university), the Chinese PHI could be extracted as translated forms in English, which can lead to a problem in surrogate generation and affect the performance evaluation procedure.

Lastly, similar to the observations in previous work [50,51], we found that the practice of requesting ChatGPT to generate responses 3 times resulted in the improvement of the overall precision, recall, and F1-score by 0.055, 0.068, and 0.064, respectively. The strategy can effectively filter out noisy outputs generated by ChatGPT. For instance, even with the inclusion of constraint statements in our prompts, we noticed that ChatGPT occasionally generated responses that violated the defined constraints, introducing undesired entity types such as “MEDICATION,” “DOSE,” or “DIAGNOSIS.” By implementing the 3-query strategy, we could successfully eliminate these cases, ensuring the quality and relevance for our guideline of the generated content. Further details regarding the limitations for this strategy and its implications are provided in the limitations section.

We explored the issue of ambiguity in the classification of fine-grained PHI at the coarse-grained level. When fine-grained PHI is misclassified, it may belong to the same coarse-grained PHI or different ones. Misclassifying it under the same coarse-grained PHI is generally more acceptable since fine-grained PHI belonging to the same coarse-grained category tends to share similar semantic characteristics. To quantify the extent of ambiguity, we calculated the micro-F-scores at the coarse-grained level, treating an incorrect recognition of a fine-grained PHI type as a true positive if it belongs to the same coarse-grained PHI type as the corresponding gold annotation. Our results (detailed in Multimedia Appendix 4) showed that intratype ambiguities are more prevalent than intertype ambiguities in “Name” and “Location” PHI types.

To gain insights into the level of ambiguity among the “Location” fine-grained types, we visualized the representations generated by M-BERT for different “Location” PHI types through t-SNE. The resulting plot (Figure 5) revealed that the clusters of “Nationality,” “Country,” and “Region” are intertwined, while the cluster of “ID” is notably separated from the others. The high ambiguity may be due to the limited number of training instances, as the number of training instances is imbalanced among PHI as illustrated in Table 2. As a result, all BERT-based models yielded better F-scores for “Country” than for “Nationality” and “Region.”

On the other hand, the “ID” PHI type had a perfect mapped recall, indicating that all of the coarse-grained PHI could be recognized by the developed models. The confusion matrix shown in Figure S1 in Multimedia Appendix 4 reveals only 1 ambiguous case where M-BERT misclassified the “Location” PHI type as the “ID” PHI type. This occurred in the sentence “She visited psychiatric clinic of 804 then but have neither regular follow-up nor fair drug compliance,” where “804” refers to the Military Taoyuan General Hospital (804 is the army number assigned to the hospital), but the model recognized it as “ID” PHI. This example demonstrates one of the most challenging cases where the model failed at recognition, even with the recall-oriented ChatGPT framework. To correctly classify the instance, the model requires additional domain knowledge. We attempted to use ChatGPT to determine whether it knows that “804 hospital” refers to the military hospital, but it replied that “804 hospital” was not a common name or official name for the hospital. However, a quick Google search revealed that the website of the hospital was the first suggested link, indicating the importance of external knowledge sources for accurate classification [56].

This section delves into how the BERT-based models use regularity and mention coverage, as outlined in the Introduction section, during their recognition process. First, we estimated the OOV rates on the test set to determine the mention coverage rate of our corpus. The estimated rates were 0.356, 0.953, 0.686, 0.876, 0.696, and 0.290 for the “Date,” “Age,” “Name,” “Location,” “Profession,” and “ID” categories, respectively. OOV was defined as a mention containing at least one word not present in the vocabulary compiled from the training data set. The inadequate performance of the DBA along with the high OOV rate suggests that the effect of the mention coverage in our corpus is limited. To analyze the effect, we re-estimated the performance of coarse-grained PHI for tokens belonging to in-vocabulary (IV) and OOV using EN-BERT, M-BERT, and TM-BERT in Figure 6. More detailed results are available in Multimedia Appendix 4.

The presented results reveal that the developed models have different levels of tolerance for OOV issues when recognizing different PHI types. Interestingly, we observed that the EN-BERT model achieved indistinguishable high F-scores for IV and OOV tokens for the “Date” PHI type. This is notable since not all Chinese characters used in the “Data” PHI are presented in the EN-BERT vocabulary. For example, the numeric value “六” (six) is not available as shown in Table 6. We believe that the presence of those characters in the extensive training data enabled EN-BERT to leverage the naming regularity of the “Date” PHI and improve its ability to identify occurrences of these entities in Chinese-only and code-mixed sentences. Figure 7 illustrates how EN-BERT’s attention mechanism can use specific patterns in text, such as “年月日,” for recognizing “Date” PHI, even when the majority of the Chinese tokens are mapped to [UNK] (unknown word) by the EN-BERT tokenizer. The model’s attention was primarily drawn to the pattern words “年” and “月,” allowing it to successfully assign labels to the respective unknown words.

However, the model’s overreliance on pattern words and strong name regularities could lead to misclassification. Figure 8 illustrates an example of the false-positive case generated by EN-BERT. The model has been trained to recognize the first likely token for “Date” PHI by considering its surrounding tokens, including previous unidentified tokens, the forward slash (/), and the final token representing the end of the “Date” PHI. However, this mechanism caused the model to incorrectly extract the text “5/4/3” as “Date” PHI. The third AHV displayed in Figure 4 illustrates that when the previous tokens are not unknown, the model focuses more on the preceding token (a number) and gives less attention to the pattern tokens (“/”), which helps prevent the occurrence of false positives.

On the other hand, the “Age,” “Location,” and “Profession” PHI types exhibited higher IV F-scores compared to their OOV scores. Specifically, the “Age” PHI type demonstrated an almost perfect IV F-score and slightly lower OOV score across all 3 models. After examining the outcomes of the developed models, we arrived at a conclusion similar to that of the “Date” PHI. The models could distinguish the “Age” PHI type in various code-mixing–level sentences because of the consistent naming pattern. There were only a few cases in which the developed models failed. For instance, in the sentence, “occupation: 會計18y->美容師3y->保險電話銷售人員->未上市股票電話銷售員->賣雞排 (2223.3-2224.9),” only CH-BERT was able to correctly avoid classifying “18” as “Age” PHI, while the other BERT-based models and the developed ChatGPT framework recognized “18” as “Age” PHI.

In contrast, the “Location” and “Profession” PHI types exhibited lower IV F-scores and apparently lower OOV F-scores. To investigate the discrepancies, we probed the results of the developed model by applying input reduction. We took the input sentence “travel history: [PHI] last week” as an example to summarize the observations as follows. First, we noticed that for EN-BERT, the presence of unknown tokens in the input sentence resulted in heavy reliance on the regularity of entity naming and contextual patterns for effective generalization over unseen mentions or mentions that contain unknown words. In the left part of Figure 9, the AHV represents the reduced interpretation of the 2 “City” PHI inputs “台北” (Taipei) and “恆春” (Hengchun) that contain unknown words. From the view, we can see that the preceding token, “travel,” provides a strong indication for EN-BERT to recognize “City” PHI. The word “台北” (“[UNK]北” for EN-BERT) was considered as “Hospital” PHI, indicating that the model learned to memorize the most frequent label assigned to the IV word, such as “署/[UNK]北” hospitals, in the training set.

Second, in addition to the naming regularity and contextual patterns, CH-BERT and M-BERT can leverage their representation for the given token sequence to assign the corresponding labels better. This is evident from the right part of Figure 5 in which we can see after fine-tuning, M-BERT was able to improve its representation of “Hospital” PHI entities, such as “ntuh,” “台大” (National Taiwan University Hospital), and “署北” (Taipei Hospital), as well as “Date” PHI entities, such as “九月” (September) and “mk87,” by clustering them in closer proximity to each other. Despite being OOV entities, OOV PHI, such as “mk99,” “菜園” (vegetable garden), “軍醫院” (military hospital), and “小診所” (small clinic), were still represented in close proximity to the corresponding IV PHI, such as “mk87,” “clothing factory,” and “療養院” (nursing home). However, “Profession” PHI and some of the fine-grained “Location” PHI types, such as “Generic location,” lacked the regularity of their naming conventions, which posed a challenge for accurate recognition. Moreover, informal language, typographical errors, ambiguity issues, and limited training instances for fine-grained PHI, such as “Region” and “Market,” further complicate the task of correctly identifying these OOV PHI mentions.

To mitigate the aforementioned OOV issue, we highlight several promising research directions. The first direction is data augmentation. During the pretraining phases, BERT-based models can benefit from automatically generating synthetic training examples that include PHI. This can be accomplished by leveraging LLMs to create data of reasonable quality [57]. This approach, similar to data augmentation, should help improve the model’s robustness when handling different contexts and OOV PHI. The second direction is vocabulary expansion. For specific categories like “Location” and “Profession,” where we observed lower F-scores on OOV PHI, traditional dictionary expansion methods [58] can be applied to augment BERT’s vocabulary. This approach empowers the model to recognize and process OOV and UNK words more effectively. The third direction is knowledge graph integration [59]. This approach enriches the contextual information available to the model, providing additional context and semantics to assist in recognizing OOV or UNK PHI and thereby improving overall performance.

While our study has made significant contributions to the field of clinical text deidentification in code-mixed languages, it is important to acknowledge some limitations. First, despite the successes achieved, M-BERT, our best-performing model, has some limitations. Notably, it exhibits intratype ambiguities in certain PHI, such as “Location,” likely due to the limited number of training instances. Second, M-BERT yields decreased performance for OOV terms compared to IV terms in specific PHI categories, such as “Profession” and “Location.” Third, our study was based on a Chinese-English code-mixed corpus from Taiwan, which has its unique writing conventions and style. As a result, the fine-tuned models may not be entirely applicable to other bilingual or multilingual code-mixed corpora from different countries.

Another notable limitation is related to the normalization of mixed sentences into their dominant word-usage language based on the CMI. In the training process for TM-BERT, we relied on an existing translation model as a means to normalize code-mixed text. While this strategy was applied to alleviate the code-mixing challenge, it is crucial to note that translation errors because of the limited context can introduce syntactic and semantic issues during fine-tuning, which can, in turn, impact the model’s overall performance. This limitation underscores the need for further research to develop more sophisticated and accurate code-mixed text normalization techniques, as explored previously [60].

In the context of the proposed ChatGPT-based deidentification framework, we acknowledge that the framework’s performance is sensitive to several factors, including the crafted prompts, the number of query times applied for the majority voting strategy, and the specific model version used. For example, we observed that when using the recently released model (version from September 27, 2023) for recognizing PHI, it experienced a higher frequency of failures due to its reluctance to process sensitive content, as compared to the version we employed in our experiments. To address this limitation, it is needed to incorporate new specific constraints, such as “None of your responses will contain ‘I'm sorry,’ ‘I apologize,’ ‘I'm sorry, but I can’t assist with that request,’ or similar.” These constraints were introduced to mitigate potential issues, but they also shed light on another limitation. As ChatGPT is subject to continuous updates by OpenAI, the specific model version we used for our study may not always reflect the latest advancements in performance. This dynamic nature of the model versions could impact the generalizability of our results and the application of the framework in evolving contexts. We acknowledge this as a potential limitation, and it is an area where ongoing research and adaptation will be essential to keep pace with model improvements and changes.

Lastly, while our decision to use the prompt 3 times is consistent with prior methodologies [50,51], we acknowledge that the specific number of repetitions can be subject to experimentation and may depend on the context and objectives of the task. Future research could delve into the selection of repetitions and its influence on result distribution to gain a deeper understanding of this aspect.

We developed the first-ever code-mixed clinical deidentification corpus, which exhibits a higher CMIC value compared to other known code-mixed data sets, suggesting that Chinese-English code-mixed clinical writing style is extensively used in the presented data set. We found that different PHI types had preference in their occurrence within the sentence categories, with “Date” and “Age” PHI types more common in English-only sentences, and “Name” and “Profession” PHI types more common in code-mixed sentences.

Our hypothesis that PLMs rely on naming regularity to recognize PHI was supported by our experimental results. Additionally, the results indicated that PLMs like M-BERT also use their learned representations to enhance their classification performance. We found that the fine-tuned M-BERT models outperformed the other approaches in most PHI types. We also observed that fine-tuning with code-mixed training instances is critical for significantly improving performance on the code-mixed data set. This finding emphasizes the significance of incorporating code-mixed training instances into the model’s training data. Our analysis revealed that TM-BERT yielded lower performance, suggesting that machine translation is not required for M-BERT in addressing the challenges posed by code mixing owing to the error-prone nature of machine-translated sentences.

We conclude that the LLM-based deidentification method is a feasible and appealing approach, which requires little to no machine learning expertise, and can be controlled and enhanced through natural language prompts. The experience of engineering our prompt also emphasizes the crucial role of carefully crafted prompts to avoid unwanted output. Further research could explore the augmentation of PLMs and LLMs with external knowledge to improve the strength in recognizing rare PHI and could incorporate experiments using the recently released GPT-4, which has been reported to demonstrate significant improvements in many tasks. However, the use of such a method in the hospital setting requires careful consideration of data security and privacy concerns.