Electronic medical record standardized conversion method and system based on Southern Fujian dialect feature recognition

Abstract

The invention provides a southern Fujian dialect feature recognition-based electronic medical record standardized conversion method and system. The method comprises the following steps of constructing a southern Fujian dialect acoustic-semantic joint modeling framework; establishing a southern Fujian dialect vocabulary intelligent mapping system based on a medical context; designing a symptom description standardization conversion engine; realizing a syntactic structure intelligent recombination mechanism; the whole conversion accuracy of the system meets the high-standard requirement of clinical application, and compared with a traditional mandarin recognition system and a general southern Fujian translation system, the whole conversion accuracy of the system is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical information processing technology, in particular to an electronic medical record standardization conversion method and system based on Minnan dialect feature recognition. BACKGROUND

[0002] With the rapid development of medical informatization, the electronic medical record system has become the core infrastructure of the modern medical system, bearing the key functions of recording, storing, transmitting and analyzing patient diagnosis and treatment information. According to the statistics of the National Health Commission, as of the end of 2023, the average application level of electronic medical records in tertiary hospitals in China reached 4.53, and the standardization and structuring of electronic medical records directly affect the effectiveness of downstream applications such as clinical decision support systems, medical quality control, and scientific research data mining. However, in the Minnan dialect-speaking areas, due to the widespread use of Minnan dialect by patients to describe their medical conditions, the electronic medical record system faces serious language processing obstacles.

[0003] As one of the seven major dialects of Chinese, Minnan dialect is spoken by more than 70 million people, and its phonetic system retains a large number of features of Middle Chinese, including a complete entering tone system, complex connected reading tone change rules, and unique white reading phenomenon. At the phonetic level, Minnan dialect has 7 tones and nasalized vowels, and the tone change rules are very different from those of Mandarin. The same word may produce 2-3 different tone changes in different contexts. In particular, in the medical context, the traditional Minnan dialect used by older patients retains more archaic features, such as the systematic loss of final p, t, k, m. This phonetic variation phenomenon causes the accuracy of existing Mandarin speech recognition systems to drop sharply. Tests show that the accuracy of mainstream commercial speech recognition engines in processing Minnan dialect medical speech is only 45%-60%.

[0004] The non-standardization of symptom expression is another technical challenge faced by electronic medical records in Minnan dialect-speaking areas. Minnan speakers tend to use figurative and exaggerated expressions. These colloquial expressions lack medical standardization and cannot be directly used for clinical diagnosis and data analysis. The existing simple word replacement method cannot accurately quantify the intensity level of these ambiguous descriptions, and doctors need to spend a lot of time for secondary confirmation and correction, which seriously affects the efficiency of diagnosis and treatment.

[0005] In addition, the medical context has very high requirements for conversion accuracy, and incorrect symptom records may lead to misdiagnosis or delayed treatment. Existing systems lack reliable quality control mechanisms and cannot assess the confidence of conversion results, nor can they identify and mark conversion segments that may have problems for manual review. At the same time, the lack of interactive correction mechanisms involving medical staff makes it difficult for the system to obtain high-quality feedback data for improvement. These technical defects seriously restrict the promotion and application of electronic medical record systems in Minnan dialect-speaking areas, and an integrated solution is urgently needed to break through these technical bottlenecks. SUMMARY

[0006] To overcome the deficiencies of the prior art, the present application proposes an electronic medical record standardization conversion method and system based on Minnan dialect feature recognition. The overall conversion accuracy of the system meets the high standard requirements of clinical application, and is greatly improved compared with traditional Mandarin recognition systems and general Minnan translation systems. Through the combination of BiLSTM deep network and DTW tone matching technology, the recognition difficulties of the complex tone system and the loss of coda of Minnan dialect are effectively solved. Based on massive professional medical terms and high-dimensional word vector representation, precise semantic disambiguation and vocabulary mapping are realized.

[0007] To achieve the above-mentioned purpose, the present application proposes an electronic medical record standardization conversion method based on Minnan dialect feature recognition, comprising the following steps:

[0008] S1: Constructing a Minnan dialect acoustic-semantic joint modeling framework,

[0009] S2: Establishing a Minnan dialect vocabulary intelligent mapping system based on medical context:

[0010] S3: Designing a symptom description standardization conversion engine;

[0011] S4: Realizing a syntax structure intelligent reorganization mechanism;

[0012] S5: Establishing an incremental learning and quality control mechanism.

[0013] Further, step S1 specifically includes:

[0014] S1.1: Collecting patient voice signals, extracting 13-dimensional MFCC features, 3-dimensional pitch features and 4-dimensional energy features, and forming a 20-dimensional basic acoustic feature vector;

[0015] S1.2: Constructing 7 specific tone template libraries of Minnan dialect, and calculating the similarity matrix of input voice and template through dynamic time warping (DTW) algorithm;

[0016] S1.3: Designing a three-layer BiLSTM network, the first layer has 256 neurons to process acoustic features, the second layer has 128 neurons to fuse tone similarity, and the third layer has 64 neurons to output phoneme probability distribution;

[0017] S1.4: Introducing a coda loss detection mechanism specific to Minnan dialect, and detecting the energy decay mode at the end of the syllable by setting a 0.3 second sliding window, and triggering the coda completion algorithm when the decay rate exceeds the preset threshold of 70%;

[0018] The coda completion algorithm in step S1.4 is specifically implemented as:

[0019] After detecting the loss of the final consonant, query the completion rule table according to the initial consonant type and the final vowel type; for the loss of the entering tone finals -p, -t, -k, infer through the place of articulation of the preceding initial consonant: -p is supplemented after labial initials, -t after apical initials, and -k after velar initials; for the loss of nasal finals -m, -n, -ng, judge through the tongue position of the main vowel of the final: -n is supplemented after front vowels, -ng after back vowels, and -m after rounded vowels.

[0020] Furthermore, step S2 specifically includes:

[0021] S2.1: Construct a basic dictionary containing 8,500 Minnan dialect medical terms, and each term contains four attributes: pronunciation, word class, semantic category, and frequency of use;

[0022] S2.2: Design a context-sensitive word disambiguation algorithm, extract the word vectors of the 5 words before and after the target word, and calculate the matching score with the candidate interpretations through cosine similarity;

[0023] S2.3: Implement word boundary recognition based on conditional random fields (CRF) to accurately segment verb-object inversion structures such as "shifan" (eat rice) and "xinglu" (walk) in Minnan dialect;

[0024] S2.4: Create a priority queue for medical terms. When detecting "touketong" (headache in Minnan dialect), it is preferentially mapped to the medical standard term "headache" rather than the colloquial "touxing" (headache in Mandarin);

[0025] The context-sensitive word disambiguation algorithm in step S2.2 specifically includes:

[0026] Train 200-dimensional word vectors using Word2Vec, and set the context window to 10; calculate the semantic similarity between the target word and each candidate interpretation; introduce TF-IDF weight adjustment to increase the weight of high-frequency medical words by 1.5 times; when the similarity differences of multiple candidate interpretations are less than 0.1, make a secondary judgment in combination with the词性标注 results; finally, select the interpretation with the highest comprehensive score as the mapping result.

[0027] Furthermore, step S3 specifically includes:

[0028] S3.1: Identify the symptom description patterns unique to Minnan dialect, including three types of expressions: reduplication ("tongtong" (very painful)), metaphorical ("like being pricked by a needle"), and hyperbolic ("painful to death");

[0029] S3.2: Construct symptom intensity quantization rules, map "yougou tong" (very painful) to pain scores of 7-8 levels, and "xiaoke tong" (a little painful) to 3-4 levels, and achieve precise conversion through a preset 15-level subdivision standard;

[0030] S3.3: Implement time expression normalization, convert "頂日" to "yesterday" and "後日" to "the day after tomorrow", and establish a complete time word mapping table;

[0031] S3.4: Handle the phenomenon of negative words being placed in front, identify structures such as "免食药" (don't need to take medicine) and "無愛去" (don't want to go), and adjust the word order through syntactic tree reconstruction;

[0032] The symptom intensity quantification rule in step S3.2 is specifically implemented as follows:

[0033] Establish a three-level quantification system including intensity adverbs, degree complements, and modal words; there are 12 levels of intensity adverbs such as "真" (coefficient 1.5), "有够" (coefficient 1.8), "小可" (coefficient 0.5), etc.; degree complements such as "到死" and "到暈" correspond to the highest intensity levels of 9 - 10 respectively; modal words such as "啦" and "咧" are used to fine-tune the final intensity value by ±0.5 levels; a standardized pain score of 0 - 10 is obtained through weighted calculation.

[0034] Furthermore, step S4 is specifically as follows:

[0035] S4.1: Construct a Minnan dialect syntactic template library, including 12 basic sentence pattern templates such as subject-predicate-object, verb-complement, and serial verb;

[0036] S4.2: Use dependency syntactic analysis to identify sentence components, and convert the Minnan word order to the standard Chinese word order through the shift-reduction algorithm;

[0037] S4.3: Design an algorithm for restoring omitted components, automatically supplement the omitted subject or object according to the context above, and restore the complete sentence structure;

[0038] S4.4: Implement sentence boundary detection and segmentation, and handle the common long sentences and run-on sentences in Minnan dialect;

[0039] The syntactic conversion in step S4.2 specifically includes:

[0040] First, identify 6 special sentence patterns in Minnan dialect: verb-object inversion sentence, double-object sentence, absence of ba-sentence, simplification of bei-sentence, serial verb sentence, and existential sentence; for verb-object inversion, identify the verb and object through dependency relations and reorganize them according to the standard word order; convert the "予" - sentence to the "给" - sentence and the "乎" - sentence to the "被" - sentence; when dealing with the "有" - existential sentence, convert it to "有" or "在" according to the context; complete the structure conversion through node exchange and path adjustment of the syntactic tree.

[0041] Furthermore, step S5 is specifically as follows:

[0042] S5.1: Set up a conversion confidence scoring system to calculate a comprehensive score based on three dimensions: acoustic model probability, lexical matching degree, and syntactic rationality;

[0043] S5.2: When the confidence level is below 0.7, automatically mark difficult segments and generate multiple candidate conversion results for manual selection;

[0044] S5.3: Collect manually calibrated data, and trigger model fine-tuning every 100 calibrated samples, using transfer learning to update parameters;

[0045] S5.4: Establish a conversion quality assessment index system, including three core indicators: lexical accuracy, syntactic accuracy, and semantic fidelity;

[0046] The incremental learning in step S5.3 is specifically implemented as follows:

[0047] The incremental learning in step S5.3 is specifically implemented as follows:

[0048] The elastic weight consolidation algorithm is used to prevent catastrophic forgetting. The Fisher information matrix is ​​calculated for the original model parameters as importance weights. The total loss function is:

[0049]

[0050] Among them, L task The current task loss is represented by λ = 0.01, which is the regularization coefficient, and F... i Let θ be the importance weight of the i-th parameter calculated using the Fisher information matrix. i For the current parameter, These are the original model parameters;

[0051] The diagonal elements of the Fisher information matrix are calculated using the following formula:

[0052]

[0053] Among them, E D : Expectations for dataset D;

[0054] p(y|x;θ * Given input x and parameter θ * The probability of outputting y at that time;

[0055] The partial derivative with respect to the i-th parameter;

[0056] The parameters are updated using gradient descent with a learning rate of η = 1 × 10^-5, and the training lasts for 3 epochs. The updated model is validated through A / B testing, and it is officially deployed when the performance improvement ratio is greater than 0.02.

[0057] The electronic medical record standardization and conversion system based on Minnan dialect feature recognition is applicable to the aforementioned electronic medical record standardization and conversion method based on Minnan dialect feature recognition, including:

[0058] Voice acquisition module: Equipped with an audio acquisition card with a sampling rate of 16kHz and 16-bit quantization, supporting mono and stereo input;

[0059] The acoustic processing unit is used to integrate a DSP chip to achieve real-time MFCC feature extraction with a processing latency of less than 50ms.

[0060] A vocabulary mapping engine is used to deploy a Redis in-memory database to store high-frequency vocabulary mapping tables, supporting 100,000 QPS queries.

[0061] A syntax converter for a custom syntax analyzer based on the Stanford Parser, optimized for the Minnan dialect;

[0062] The quality control module is used to implement multi-threaded concurrent processing and supports confidence evaluation of batch text.

[0063] A model updater for containerized deployment using Docker, supporting hot model updates without service interruption;

[0064] The data storage system uses MongoDB to store raw speech, converted text, and correction records, and supports horizontal scaling.

[0065] Furthermore, the vocabulary mapping engine includes:

[0066] The main dictionary storage area stores 8,500 basic entries, with each entry occupying an average of 2KB of space;

[0067] The cache manager uses the LRU algorithm to manage hot keywords, maintaining a cache hit rate of over 85%.

[0068] The dictionary features a thesaurus index with an inverted index structure, supporting fuzzy matching and homophone queries. The update synchronization mechanism utilizes a message queue to achieve real-time synchronization of dictionary updates with a latency of less than 100ms.

[0069] Furthermore, it also includes the user interface:

[0070] A web-based editor developed using Vue.js, supporting text highlighting, real-time preview, and annotation features;

[0071] Confidence levels are visualized, with high, medium, and low confidence segments marked with different colors.

[0072] The batch processing interface provides a REST API, supports JSON format data interaction, and can process up to 1MB of text in a single batch; the export function supports exporting to standardized medical record documents in Word, PDF, and HL7 CDA formats.

[0073] Compared with the prior art, the beneficial effects of the present invention are:

[0074] 1. This invention provides a standardized electronic medical record conversion method and system based on Minnan dialect feature recognition. Through a multi-dimensional confidence assessment mechanism, the system can intelligently identify potentially problematic conversion results and issue timely warnings. Low-confidence segments are highlighted with color to alert medical staff, forming a reliable quality control defense. The innovative symptom intensity quantification system converts ambiguous dialect expressions into precise medical quantitative indicators, such as accurately mapping exaggerated expressions to corresponding pain levels, avoiding misjudgments of symptom severity. The error rate of medical record recording is significantly reduced, effectively lowering the risk of misdiagnosis due to language comprehension biases and providing crucial protection for patient safety.

[0075] 2. This invention provides a method and system for the standardized conversion of electronic medical records based on Minnan dialect feature recognition. Through continuous absorption of human correction feedback, the model performance steadily improves, and the accuracy of lexical mapping and syntactic conversion is continuously enhanced. The standardization and structuring of electronic medical records are significantly improved, from an originally inconsistent quality level to a near-fully usable level.

[0076] 3. This invention provides a standardized conversion method and system for electronic medical records based on Minnan dialect feature recognition, which greatly reduces the burden of medical record entry, significantly shortens the editing time for each medical record, and the cumulative saved working hours are equivalent to increasing the output of multiple medical staff. The system supports high-concurrency batch processing and can simultaneously serve the medical record conversion needs of multiple departments, effectively alleviating the paperwork pressure on medical staff. Through automated language conversion, doctors can devote more time and energy to patient diagnosis and treatment and doctor-patient communication, improving the quality of medical services. Attached Figure Description

[0077] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0078] Figure 1 This is a schematic diagram of the system flow of the present invention. Detailed Implementation

[0079] The technical solution of the present invention will be more clearly and completely explained below with reference to the accompanying drawings and through the description of preferred embodiments of the present invention.

[0080] like Figure 1 As shown, this invention provides an electronic medical record natural language standardization conversion system based on Minnan dialect feature recognition. The system adopts a five-level serial processing architecture and realizes intelligent conversion of Minnan dialect speech to standardized electronic medical record text through the collaborative work of multiple functional modules.

[0081] The system's input end is equipped with a patient speech acquisition module, which uses a 16kHz sampling rate to digitally acquire the patient's Minnan dialect speech in real time. The speech signal first enters the S1 acoustic-semantic joint modeling module, which is the first-level processing unit of the entire system. The left sub-module of the S1 module is responsible for extracting 20-dimensional acoustic features, including 13-dimensional MFCC features, 3-dimensional pitch features, and 4-dimensional energy features. These feature vectors comprehensively characterize the time-frequency domain characteristics of the speech signal. The right sub-module performs 7-tone template matching and coda detection, specifically addressing the complex tone system and coda dropout phenomenon unique to the Minnan dialect. The core of the S1 module uses a BiLSTM deep neural network combined with the DTW dynamic time warping algorithm. The BiLSTM network is designed with a three-layer structure of 256-128-64 to achieve accurate recognition of Minnan dialect speech and output preliminary text transcription results.

[0082] After speech recognition, the text data flows to the S2 intelligent vocabulary mapping system, which is the second-level processing unit. The S2 module integrates Graph Neural Network (GNN) and Conditional Random Field (CRF) technologies. The left submodule maintains a professional dictionary containing 8547 Minnan dialect medical terms, representing semantic information through 200-dimensional word vectors. The right submodule is responsible for context disambiguation and TF-IDF weighted calculation. When encountering polysemous words, it determines the most accurate semantic mapping by analyzing the preceding and following context and word frequency weights. This module accurately converts Minnan dialect vocabulary into corresponding standard Mandarin medical terms, achieving a vocabulary mapping accuracy of 92.5%.

[0083] The S3 symptom description standardization engine, as the third-level processing unit, is specifically designed to standardize symptom expressions in medical scenarios. This module establishes a 15-level fine-grained quantification system. The left sub-module identifies three special expression patterns in the Minnan dialect: overlapping, metaphorical, and exaggerated expressions. The right sub-module applies intensity quantification rules and time expression standards to convert colloquial symptom descriptions into standard medical expressions. Through this module, vague dialectal symptom descriptions are transformed into precise medical terminology and quantitative indicators.

[0084] Syntactic processing is handled by the S4 syntactic structure reorganization module, the fourth-level processing unit. This module, based on dependency parsing, identifies word order features unique to the Minnan dialect, such as verb-object inversion and subject omission, and uses a syntactic tree reconstruction algorithm to adjust the Minnan word order to standard Chinese word order. Internally, the module employs an improved version of the Stanford Parser, optimized for the characteristics of medical texts, capable of handling complex sentence transformations with a syntactic accuracy rate of 95.0%.

[0085] The S5 quality control and incremental learning module is the fifth-level processing unit of the system and is crucial for ensuring continuous performance optimization. This module employs the EWC (Elastic Weight Consolidation) algorithm for incremental learning, setting a confidence threshold of 0.7 to evaluate the conversion results across three dimensions: acoustic model probability, lexical matching accuracy, and syntactic rationality. When the confidence level falls below the threshold, the system labels the corresponding segments for manual correction. The module triggers a model update every 100 corrected samples, validated through A / B testing. A new model is officially deployed when the performance improvement exceeds 2%.

[0086] The system outputs standardized medical record text, and only transformation results with a confidence level of 0.7 or higher are output. Simultaneously, the system includes a human feedback channel (shown by the dotted line in the diagram), where correction data from medical staff flows back to the S5 module, forming a closed-loop optimization mechanism. Furthermore, each major module maintains bidirectional data interaction with the auxiliary modules on the side (indicated by the dotted arrows), ensuring full utilization of information during processing.

[0087] The system achieves high-quality conversion from Minnan dialect speech to standardized electronic medical record text, effectively solving the language barrier problem faced by medical informatization in dialect areas.

[0088] As a specific implementation method, the system was deployed in the outpatient department of a tertiary hospital, covering eight departments including internal medicine, surgery, and pediatrics, processing approximately 1200 electronic medical records in the Minnan dialect daily. The system was deployed on a server configured with an Intel Xeon Gold 6248R processor (24 cores, 48 ​​threads), 256GB of memory, and four 2TB NVMe SSDs arranged in a RAID 10 array. The operating system was CentOS 7.9, and the deep learning framework used was PyTorch 1.12.0.

[0089] First, voice signal acquisition and processing are carried out. A RODE NT-USB professional microphone is configured in each consulting room and connected to the acquisition terminal through a USB 3.0 interface. The voice signal is digitized at a sampling rate of 16 kHz and a quantization precision of 16 bits and transmitted to the server in real time. In the preprocessing stage, the Webrtcvad voice activity detection algorithm is used to remove the silent segments, and spectral subtraction is used for noise reduction, with the signal-to-noise ratio increased by about 12 dB. The feature extraction module calculates one frame every 25 ms with a frame shift of 10 ms and extracts 13-dimensional MFCC features, specifically 12-order cepstral coefficients plus 1-dimensional logarithmic energy. The pitch feature is extracted by the autocorrelation function method to obtain the fundamental frequency F0 and its first-order and second-order differences, forming a 3-dimensional feature. The energy feature includes four dimensions: short-time energy, zero-crossing rate, spectral centroid, and spectral roll-off point, and finally forms a 20-dimensional acoustic feature vector.

[0090] For the 7 tones of Minnan dialect (yinping, yangping, yinshang, yangshang, yinqu, yangqu, yinru), the system pre-collected the standard pronunciations of 50 native speakers of different ages (20 - 70 years old). 100 samples were recorded for each tone, and tone templates were generated through K-means clustering. During actual recognition, the DTW algorithm is used to calculate the distance between the input voice and each template, and the distance threshold is set to 0.35 to generate a 7×T similarity matrix (T is the number of frames). The BiLSTM network adopts a three-layer structure. The first layer contains 256 LSTM units to process acoustic features with a dropout rate of 0.2; the second layer has 128 units to fuse tone similarity information; the third layer has 64 units. Finally, the probability distribution of 1547 Minnan dialect phonemes is output through a fully connected layer and softmax. The network is trained using the CTC loss function with an initial learning rate of 0.001, the Adam optimizer, a batch size of 32, and trained for 50 epochs on a dataset containing 200 hours of labeled speech. The final phoneme recognition accuracy reaches 87.3%.

[0091] For the detection of coda deletion, a 0.3-second sliding window with a step size of 0.05 seconds is used to calculate the slope of the energy envelope within the window. When the detected energy decay rate exceeds 70% and the duration is less than 0.1 second, it is determined as coda deletion. For example, the entering tone coda -k in "shifan" often drops. After the system detects a sudden drop in the energy at the end of the character "shi", it infers that the -k coda should be supplemented based on the initial consonant s- (voiceless fricative) and restores it to the complete syllable [sik8]. The actual measurement shows that the coda completion accuracy reaches 82.5%, among which the entering tone coda completion accuracy is the highest at 85.2%, and the nasal coda is 79.8%.

[0092] The basic dictionary of the vocabulary mapping system is constructed by crawling resources such as Minnan Wikipedia and the Common Minnan Words Dictionary of the Ministry of Education, and combined with manual annotation of 3-year historical medical records in the hospital. A total of 8,547 Minnan medical-related entries are included. The storage structure of each entry includes: Minnan characters (such as "跤頭趺"), pinyin, word class (noun), semantic category (body part), standard word mapping (knee), and usage frequency (statistically based on historical medical records, the word appears 423 times per month on average). The word vectors are trained using the Skip-gram model of Word2Vec, with a window size of 10, a vector dimension of 200, and a negative sampling number of 5. They are trained for 100 rounds on a Minnan medical corpus of 5 million words. The accuracy of the semantic similarity evaluation of the final word vectors (tested by 500 pairs of synonyms constructed manually) reaches 91.2%.

[0093] When implementing context disambiguation, take the word "燒" as an example. In Minnan language, this word can mean both "fever" and "sense of burning". The system extracts the word vectors of the 5 words before and after. If words such as "溫度" and "38度" are detected in the previous text, the cosine similarity with "發燒" is calculated to be 0.83, and the similarity with "灼熱" is 0.41. Therefore, it is mapped to "fever". At the same time, the TF-IDF weight is introduced. The IDF value of "發燒" in the medical corpus is 2.3, and that of "灼熱" is 3.8. The final weighted scores are 1.909 and 1.558 respectively, and "fever" is confirmed as the choice. In actual tests, the accuracy of polysemous word disambiguation reaches 88.7%.

[0094] In the standardized processing of symptom descriptions, the system identifies three types of special expression patterns. After the overlapping pattern such as "痛痛" is identified, it is mapped to "continuous pain" through rules; the metaphorical pattern "像針拄" (like a needle prick) is converted to "stabbing pain"; the hyperbolic pattern "痛甲要死" calculates the intensity value based on modal particles and adverbs of degree. When quantifying specifically, the coefficient of "真" in "真痛" is assigned 1.5, and the basic pain level is 5. The calculated result is 7.5 levels; the coefficient of "有够" in "有够痛" is 1.8, and the result is 9 levels; the coefficient of "小可痛" is 0.5, and the result is 2.5 levels. The system has established a quantification dictionary containing 43 adverbs of degree and 28 modal words. In the test of 1,000 symptom descriptions, the conversion accuracy rate evaluated by doctors is 92.1%, and the accuracy rate of pain intensity judgment reaches 89.5%.

[0095] Confidence evaluation uses three-dimensional scoring: the output probability of the acoustic model (weight 0.4), the lexical matching score (weight 0.3), and the syntactic rationality score (weight 0.3). When "跤頭趺痛三工" is translated to "knee pain for three days", the average probability of the acoustic model is 0.82, the lexical matching degree is 0.91 (3 out of 4 words have high confidence), the syntactic score is 0.88, and the comprehensive confidence is 0.82×0.4 + 0.91×0.3 + 0.88×0.3 = 0.865. The system sets 0.7 as the threshold. Fragments below this value are highlighted in yellow, and those below 0.5 are marked in red and three candidates are provided.

[0096] Incremental learning is implemented using the EWC algorithm. First, the Fisher information matrix is calculated on the original model, and 5000 data samples are sampled to estimate the importance of each parameter. After collecting 100 manually corrected samples, a fine-tuning dataset is constructed. Using the AdamW optimizer, with a learning rate of 1e-5 and a weight decay of 0.01, training is carried out for 3 epochs. After the first incremental learning, the lexical mapping accuracy rate increases from 88.7% to 90.1%, the second time it increases to 91.3%, and the third time it reaches 91.8%, and the improvement amplitude gradually converges. Through A / B testing, 10% of the traffic is directed to the new model, monitored for 7 days, and after confirming that the performance improvement exceeds 2% and there are no abnormalities, full-scale deployment is carried out.

[0097] Statistical data from actual deployment and operation for 6 months shows that the system processes approximately 180 hours of voice per day on average, with a text volume of approximately 1.2 million words. The overall conversion accuracy rate reaches 91.5%, including a voice recognition accuracy rate of 90.7%, a lexical mapping accuracy rate of 92.5%, and a syntactic normalization accuracy rate of 95.0%. The doctor satisfaction survey shows that 93.8% of doctors believe that the system has significantly improved the medical record entry efficiency, saving an average of 3.2 minutes of editing time per medical record. Compared with the manually transcribed medical records before deployment, the standardization level has increased by 47%, and the availability rate for subsequent clinical decision support systems has increased from 61% to 94%. The average system response delay is 187ms, the peak QPS reaches 12000, the CPU utilization rate remains at about 45%, and the memory occupancy is 18GB, fully meeting the business needs of the hospital.

[0098] The comparative experiment selected three benchmark systems: a pure Mandarin speech recognition system (iFlytek), a general Minnan dialect translation system (Academia Sinica, Taiwan), and a simple dictionary replacement system (self-built). On the same 1000 test samples, the Mandarin system achieved an accuracy of only 62.3%, mainly due to its inability to recognize Minnan dialect; the general translation system achieved an accuracy of 71.5%, lacking adaptation for the medical field; and the dictionary replacement system achieved an accuracy of 78.2%, unable to handle contextual and syntactic issues. Our system, with an accuracy of 91.5%, significantly outperformed all benchmark systems and achieved the best performance in the three dimensions of medical terminology standardization, semantic accuracy, and sentence fluency.

[0099] The above-described specific embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications, substitutions, and improvements made by those skilled in the art to the technical solutions of the present invention based on the provided textual description and drawings, without departing from the design concept and spirit of the present invention, should all fall within the scope of protection of the present invention. The scope of protection of the present invention is determined by the claims.

Claims

1. A method for standardizing and converting electronic medical records based on Minnan dialect feature recognition, characterized in that, Includes the following steps: S1: Constructing a joint acoustic-semantic modeling framework for the Minnan dialect; S2: Establish an intelligent mapping system for Minnan dialect vocabulary based on medical context; S3: Design a standardized conversion engine for symptom descriptions; S4: Implements an intelligent syntactic structure reorganization mechanism; S5: Establish an incremental learning and quality control mechanism.

2. The electronic medical record standardization conversion method based on Minnan dialect feature recognition according to claim 1, characterized in that, Step S1 specifically includes: S1.1: Collect the patient's speech signal, extract 13-dimensional MFCC features, 3-dimensional pitch features and 4-dimensional energy features to form a 20-dimensional basic acoustic feature vector; S1.2: Construct a library of 7 tone templates unique to the Minnan dialect, and calculate the similarity matrix between the input speech and the templates using the Dynamic Time Warping (DTW) algorithm; S1.3: Design a three-layer BiLSTM network. The first layer has 256 neurons to process acoustic features, the second layer has 128 neurons to fuse tone similarity, and the third layer has 64 neurons to output phoneme probability distribution. S1.4: Introduces a unique coda-dropping detection mechanism for Minnan dialect. By setting a 0.3-second sliding window to detect the energy decay mode at the end of a syllable, the coda-filling algorithm is triggered when the decay rate exceeds a preset threshold of 70%. The specific implementation of the final consonant completion algorithm in step S1.4 is as follows: After detecting the loss of a final consonant, the completion rule table is consulted based on the type of initial consonant and final consonant. For the loss of entering tone final consonants p, t, and k, the rule is inferred from the articulation position of the preceding initial consonant: p is added after labial initial consonants, t after alveolar initial consonants, and k after velar initial consonants. For the loss of nasal final consonants m, n, and ng, the rule is determined from the tongue position of the main vowel of the final consonant: n is added after front vowels, ng after back vowels, and m after rounded vowels.

3. The electronic medical record standardization conversion method based on Minnan dialect feature recognition according to claim 1, characterized in that, Step S2 specifically includes: S2.1: Construct a basic dictionary containing 8,500 medical terms in the Minnan dialect, with each term containing four attributes: pronunciation, part of speech, semantic category, and frequency of use; S2.2: Design a context-sensitive word disambiguation algorithm to extract the word vectors of the five words before and after the target word and calculate the matching score with the candidate definition using cosine similarity. S2.3: Implement lexical boundary recognition based on conditional random fields to solve the problem of accurate segmentation of verb-object inversion structures in Minnan dialect; S2.4: Create a medical terminology priority queue. When "headache" is detected, it is preferentially mapped to the medical standard term "headache" rather than the colloquial "headache". The context-sensitive word disambiguation algorithm in step S2.2 specifically includes: 200-dimensional word vectors were trained using Word2Vec, with the context window set to 10. Calculate the semantic similarity between the target word and each candidate definition; By introducing TF-IDF weight adjustment, the weight of high-frequency medical terms increased by 1.5 times; When multiple candidate definitions have a similarity difference of less than 0.1, a secondary judgment is made based on the part-of-speech tagging results. The interpretation with the highest overall score was ultimately selected as the mapping result.

4. The electronic medical record standardization conversion method based on Minnan dialect feature recognition according to claim 1, characterized in that, Step S3 specifically includes: S3.1: Identify the symptom description patterns unique to Minnan dialect, including three types of expressions: reduplication, metaphor, and hyperbole; S3.2: Construct symptom intensity quantification rules, map "severely painful" to pain scores of level 7 - 8, and "slightly painful" to level 3 - 4, and achieve precise conversion through a preset 15 - level subdivision standard; S3.3: Implement normalization of time expressions, convert "the day before yesterday" to "yesterday", and "the day after tomorrow" to "the day after tomorrow", and establish a complete mapping table of time words; S3.4: Handle the phenomenon of negative words being placed in front, identify the structure, and adjust the word order through syntactic tree reconstruction; The symptom intensity quantification rules in step S3.2 are specifically implemented as follows: Establish a three - level quantification system including intensifying adverbs, degree complements, and modal words; The intensifying adverbs include 12 levels; the degree complements respectively correspond to the highest intensity of level 9 - 10; The modal words are used to fine - tune the final intensity value by ±0.5 levels; Obtain a standardized pain score of 0 - 10 through weighted calculation.

5. The electronic medical record standardization conversion method based on Minnan dialect feature recognition according to claim 1, characterized in that, Step S4 is specifically as follows: S4.1: Construct a Minnan dialect syntactic template library, including 12 basic sentence pattern templates; S4.2: Use dependency syntactic analysis to identify sentence components, and convert the Minnan word order to the standard Chinese word order through the shift - reduction algorithm; S4.3: Design an algorithm for restoring omitted components, automatically supplement the omitted subject or object according to the context above, and restore the complete sentence structure; S4.4: Implement sentence boundary detection and segmentation, and handle the common long - sentence and run - on - sentence phenomena in Minnan dialect; The syntactic conversion in step S4.2 specifically includes: First, identify 6 special sentence patterns in Minnan dialect: verb - object inversion sentence, double - object sentence, absence of the "ba" sentence, simplification of the "bei" sentence, serial verb construction sentence, and existential sentence; For verb - object inversion, identify the verb and object through dependency relations and reorganize them according to the standard word order; Convert the "yu" sentence to the "gei" sentence, and the "hu" sentence to the "bei" sentence; when dealing with the "you" existential sentence, convert it to "you" or "zai" according to the context; complete the structure conversion through node exchange and path adjustment of the syntactic tree.

6. The electronic medical record standardization conversion method based on Minnan dialect feature recognition according to claim 1, characterized in that, Step S5 is specifically as follows: S5.1: Set up a conversion confidence scoring system, calculate the comprehensive score based on three dimensions: acoustic model probability, lexical matching degree, and syntactic rationality; S5.2: When the confidence is lower than 0.7, automatically mark the difficult fragments and generate multiple candidate conversion results for manual selection; S5.3: Collect manually corrected data, trigger model fine - tuning every time 100 correction samples are accumulated, and update the parameters using transfer learning; S5.4: Establish a conversion quality evaluation index system, including three core indicators: lexical accuracy, syntactic correctness, and semantic fidelity; The incremental learning in step S5.3 is specifically implemented as follows: Adopt the elastic weight consolidation algorithm to prevent catastrophic forgetting, calculate the Fisher information matrix as the importance weight for the original model parameters, and the total loss function is: Among them, L task The current task loss is represented by λ = 0.01, which is the regularization coefficient, and F... i Let θ be the importance weight of the i-th parameter calculated using the Fisher information matrix. i For the current parameter, These are the original model parameters; The diagonal elements of the Fisher information matrix are calculated by the following formula: Among them, E D : Expectations for dataset D; p(y|x;θ * Given input x and parameter θ * The probability of outputting y at that time; The partial derivative with respect to the i-th parameter; The parameter update uses the gradient descent method, the learning rate η = 1×10^-5, and train for 3 epochs; the updated model is verified through A / B testing, and when the performance improvement ratio is greater than 0.02, it is officially deployed.

7. A system for standardizing and converting electronic medical records based on Minnan dialect feature recognition, applicable to the method for standardizing and converting electronic medical records based on Minnan dialect feature recognition as described in any one of claims 1-6, characterized in that, Include: Voice acquisition module: Configure an audio acquisition card with a sampling rate of 16 kHz and 16 - bit quantization, supporting mono and stereo inputs; The acoustic processing unit is used to integrate a DSP chip to achieve real-time MFCC feature extraction with a processing latency of less than 50ms. A vocabulary mapping engine is used to deploy a Redis in-memory database to store high-frequency vocabulary mapping tables, supporting 100,000 QPS queries. A syntax converter for a custom syntax analyzer based on the Stanford Parser, optimized for the Minnan dialect; The quality control module is used to implement multi-threaded concurrent processing and supports confidence evaluation of batch text. A model updater for containerized deployment using Docker, supporting hot model updates without service interruption; The data storage system uses MongoDB to store raw speech, converted text, and correction records, and supports horizontal scaling.

8. The electronic medical record standardization and conversion system based on Minnan dialect feature recognition according to claim 7, characterized in that, The vocabulary mapping engine includes: The main dictionary storage area stores 8,500 basic entries, with each entry occupying an average of 2KB of space; The cache manager uses the LRU algorithm to manage hot keywords, maintaining a cache hit rate of over 85%. The dictionary features a thesaurus index with an inverted index structure, supporting fuzzy matching and homophone queries. The update synchronization mechanism utilizes a message queue to achieve real-time synchronization of dictionary updates with a latency of less than 100ms.

9. The electronic medical record standardization and conversion system based on Minnan dialect feature recognition according to claim 7, characterized in that, It also includes the user interface: A web-based editor developed using Vue.js, supporting text highlighting, real-time preview, and annotation features; Confidence levels are visualized, with high, medium, and low confidence segments marked with different colors. The batch processing interface provides a REST API, supports JSON format data interaction, and can process up to 1MB of text at a time. The export function supports exporting standardized medical record documents in Word, PDF, and HL7 CDA formats.

Images