Traceable sleep disorder assisted analysis method and apparatus based on pointer anchoring evidence

Abstract

The present application provides a kind of traceable sleep disorder auxiliary analysis method and device based on pointer anchoring evidence, belongs to data processing technical field, by introducing and initializing explicit evidence pointer library, so that the evidence atom field in the output structured analysis result must belong to the pointer library, the freedom degree of model generation evidence is limited.Simultaneously, the parameter training of two stages of sleep disorder auxiliary analysis model makes the accuracy of internal multi-modal feature extraction of sleep disorder auxiliary analysis model greatly improve when processing high-density, multi-page mixed PSG report data of text and image, avoids the omission and misreading of basic fact data.At the same time, reduce the risk of unfounded statement downstream propagation, and improve the traceability and auditability of structured output.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a traceable sleep disorder auxiliary analysis method and apparatus based on pointer-anchored evidence. Background Technology

[0002] With the rapid development of Large Language Models (LLM) and Multimodal Large Language Models (MLM), the application of human intervention in the medical field is evolving towards automated clinical decision support. In sleep medicine, polysomnography (PSG) reports are crucial objective evidence for assessing sleep-related breathing disorders such as obstructive sleep apnea. For sleep disorders like insomnia, a comprehensive analysis is usually required, combining the subject's chief complaint, medical history, scale results, and other clinical documents. A complete PSG report typically includes a multi-page, high-density mix of text and graphics, including a summary of the subject's sleep structure, respiratory event indicators, blood oxygenation trends, and postural correlation charts. When analyzing the report, clinicians not only need to review the various objective indicators but also need to make comprehensive inferences based on the subject's chief complaint, medical history, and other clinical documents.

[0003] Currently, medical document processing solutions based on existing large-scale modeling technologies are mainly divided into two categories: text generation solutions and general visual understanding solutions. However, when applying existing technologies directly to the clinical analysis of sleep disorders, the following significant defects and shortcomings exist: First, the reading stability of indicators in complex, long medical documents is poor, making them prone to omissions and misinterpretations. Existing general-purpose multimodal models are highly susceptible to visual distraction when faced with PSG reports spanning multiple pages and containing dense tables and trend graphs. Due to a lack of training for reading alignment with high-density charts, models often struggle to accurately extract key physiological indicator values, directly leading to biases in the underlying facts of subsequent analysis and inference.

[0004] Second, the lack of factual traceability in the analysis results easily leads to model illusion. Existing decision support models mostly output analytical conclusions and recommendations in the form of free text or long sentences. Even when some systems attempt to provide evidence, the cited evidence is often vague or freely generated. This highly free generation method leads to the system frequently outputting misleading judgments without direct evidence support. For clinicians and medical regulatory agencies, facing massive amounts of generated text, it is difficult to consistently pinpoint each analytical statement to the original report fields or clinical text fragments, reducing the verifiability and auditability of the conclusions and potentially posing medical safety risks.

[0005] In summary, how to provide a method that can accurately read complex multi-page charts, anchor analytical statements to original evidence, and thus output structured and traceable auxiliary analysis results for sleep disorders has become a pressing technical challenge in the field of clinical medical artificial intelligence. Summary of the Invention

[0006] This invention provides a traceable sleep disorder auxiliary analysis method and apparatus based on pointer-anchored evidence to address the deficiencies in the prior art. It significantly improves the accuracy of multimodal feature extraction in the sleep disorder auxiliary analysis model when processing high-density, multi-page mixed text and image PSG report data, avoids the omission and misinterpretation of basic factual data, reduces the risk of unfounded claims spreading downstream, and improves the traceability and auditability of structured output.

[0007] This invention provides a traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence, comprising the following steps.

[0008] Obtain the polysomnography report of the target subject and the corresponding clinical documents of the polysomnography report; Initialize the explicit evidence pointer library; The polysomnography report, the corresponding clinical document, and the explicit evidence pointer library are input into the sleep disorder auxiliary analysis model for parsing, and the structured analysis results output by the sleep disorder auxiliary analysis model include at least the analysis result field, evidence atom field, explanatory field, and adjustment suggestion field. The evidence atom field contains at least one evidence atom to support the analysis result field, and the evidence pointer field of each evidence atom is taken from the evidence pointer in the explicit evidence pointer library. The sleep disorder auxiliary analysis model is trained on the report reading alignment model using the first input sample and its corresponding explicit evidence pointer library as input and the structured analysis label as the supervision signal. The report reading alignment model is trained on the polysomnography report sample and the sleep index label corresponding to the polysomnography report sample. The first input sample includes the polysomnography report sample and the clinical document corresponding to the polysomnography report sample.

[0009] According to the present invention, a traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence is provided, wherein the evidence pointers in the explicit evidence pointer library include: Reporting indicator pointers are used to identify standardized indicators in the polysomnography report, and clinical text pointers are used to identify standardized anchor points in the clinical text.

[0010] According to the present invention, a traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence is provided, wherein the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence further includes: Based on the normalization results of the indicator names, indicator values, units, page positions, table fields, and chart areas in the polysomnography report, the report indicator pointers are generated. Anchor points are extracted from the chief complaint, present illness, symptoms, and medical history fragments in the clinical file corresponding to the polysomnography report to generate the clinical text pointer; The explicit evidence pointer library is constructed based on the reported indicator pointers and the clinical text pointers.

[0011] According to the present invention, a traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence is provided, wherein the evidence atom includes at least the following information: The Fact Claim field is used to describe the factual information claimed by the evidence atom; The source type field describes the source type of the evidence atom; The Evidence Pointer field describes the evidence pointer referenced by the evidence atom.

[0012] According to the present invention, a traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence is provided, wherein the description field includes at least the following information: The description item's type information, encoding information, target information, evidence pointer field referenced by the description item, and description text. The type information of the description item is one of the following types: Missing information, supporting information, conflicting information, next steps suggestions, and conservative output information; The conservative output information is used to indicate that due to insufficient evidence, a judgment cannot be made and manual review is recommended.

[0013] According to the present invention, a traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence is provided, wherein the number of evidence atoms contained in the evidence atom field is less than or equal to a preset number threshold.

[0014] According to the present invention, a traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence is provided, wherein the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence further includes: The structured analysis results are verified using preset verification items; If the structured analysis results pass the verification, the structured analysis results are output. If the structured analysis result fails the verification, the output of the structured analysis result is canceled, and a conservative structured analysis result, information on missing items in the conservative structured analysis result, and result prompt information are generated. The result prompts include prompts for insufficient evidence or prompts for manual review; The preset verification items include one or more of the following: The following checks are performed: structured format integrity verification; whether the evidence atom field contains evidence atoms; whether the evidence pointer belongs to the pointer attribution of the explicit evidence pointer library; the number of evidence atoms corresponding to the evidence atoms contained in the evidence atom field; consistency verification of the structured analysis result content; information missing verification; whether to output conservative structured analysis results when the evidence atom field does not contain evidence atoms; and whether there is a conflict between the missing information in the description field and the analysis result field.

[0015] According to the present invention, a traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence is provided, wherein the report reading alignment model is trained through the following steps: The polysomnography report samples and pre-built structured extraction constraint prompts are input into the initial model to extract indicators, and the extraction results corresponding to each indicator are obtained. Using the sleep index labels corresponding to the polysomnography report samples as a reference, a rule-based scorer is used to evaluate each extraction result to obtain a score value for each extraction result. Based on the scores of all the extraction results, the initial model is optimized using a pre-selected model optimization algorithm until the optimization conditions are met; The structured extraction constraints are used to constrain the content contained in the extraction results output by the initial model.

[0016] According to the present invention, a traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence is provided, wherein the sleep disorder auxiliary analysis model is trained through the following steps: The polysomnography report sample in the first input sample is input into the visual encoder for feature extraction to obtain a visual feature map sequence. The visual feature map sequence is converted into a visual word sequence using a visual-language fusion module; The clinical documents corresponding to the polysomnography report samples and the explicit evidence pointer library are input into a large language model for mapping to obtain a text word sequence; The visual word sequence and the text word sequence are concatenated and input into the large language model for prediction to obtain structured analysis prediction; Based on the cross-entropy loss of the structured analysis prediction and the structured analysis label corresponding to the first input sample, update at least one of the first trainable parameter, the second trainable parameter, and the third trainable parameter until the training conditions are met. The first trainable parameter is the trainable parameter of the visual encoder, the second trainable parameter is the trainable parameter of the large language model, and the third trainable parameter is the trainable parameter of the visual-language fusion module.

[0017] The present invention also provides a traceable sleep disorder auxiliary analysis device based on pointer-anchored evidence, comprising the following modules: The acquisition module is used to acquire the polysomnography report of the target subject and the corresponding clinical documents of the polysomnography report; The initialization module is used to initialize the explicit evidence pointer library; The parsing module is used to input the polysomnography report, the corresponding clinical document of the polysomnography report and the explicit evidence pointer library into the sleep disorder auxiliary analysis model for parsing, and to obtain the structured analysis results output by the sleep disorder auxiliary analysis model, which includes at least the analysis result field, evidence atom field, explanatory field and adjustment suggestion field. The evidence atom field contains at least one evidence atom to support the analysis result field, and the evidence pointer field of each evidence atom is taken from the evidence pointer in the explicit evidence pointer library. The sleep disorder auxiliary analysis model is trained on the report reading alignment model using the first input sample and its corresponding explicit evidence pointer library as input and the structured analysis label as the supervision signal. The report reading alignment model is trained on the polysomnography report sample and the sleep index label corresponding to the polysomnography report sample. The first input sample includes the polysomnography report sample and the clinical document corresponding to the polysomnography report sample.

[0018] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence as described above.

[0019] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence as described above.

[0020] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence as described above.

[0021] This invention provides a traceable sleep disorder auxiliary analysis method and apparatus based on pointer-anchored evidence. By introducing and initializing an explicit evidence pointer library, the atomic fields of evidence in the output structured analysis results must be mapped to this pointer library, limiting the degree of freedom of the model in generating evidence and effectively suppressing the illusion of the model piling up false information. Simultaneously, the two-stage parameter training of the sleep disorder auxiliary analysis model significantly improves the accuracy of its internal multimodal feature extraction when processing high-density, multi-page mixed text and image PSG report data, avoiding the omission and misinterpretation of basic factual data. Furthermore, it reduces the risk of unfounded claims propagating downstream and improves the traceability and auditability of the structured output. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is one of the flowcharts of the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence provided by the present invention.

[0024] Figure 2 This is the second flowchart of the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence provided by the present invention.

[0025] Figure 3 This is a schematic diagram of the training process of the report reading alignment model provided by the present invention.

[0026] Figure 4 This is a schematic diagram of the training process of the sleep disorder auxiliary analysis model provided by the present invention.

[0027] Figure 5 This is a schematic diagram illustrating the training principle of the sleep disorder auxiliary analysis model provided by the present invention.

[0028] Figure 6 This is a schematic block diagram of the traceable sleep disorder auxiliary analysis device based on pointer-anchored evidence provided by the present invention.

[0029] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0031] This application provides a traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence. The executing entity can be an electronic device with data processing and model reasoning capabilities, such as, but not limited to, a local server, a cloud server cluster, an in-hospital medical auxiliary diagnostic workstation system, an edge computing node, or a dedicated data processing device integrated into a hospital management information system or an electronic medical record system.

[0032] The following is combined Figures 1 to 7 The present invention describes a traceable sleep disorder auxiliary analysis method and apparatus based on pointer-anchored evidence.

[0033] Figure 1 This is one of the flowcharts of the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence provided by the present invention, such as... Figure 1 As shown, the method includes the following: Step 101: Obtain the polysomnography report of the target subject and the corresponding clinical documents.

[0034] Polysomnography reports are typically datasets that exist as multi-page images, scanned archives, or mixed image and text formats. They naturally contain complex layout information such as summaries of the subject's sleep structure, high-density respiratory event indicators, blood oxygenation-related indicators, body position-related information, and graphical trends.

[0035] Clinical documents represent textual information generated by doctors during consultations or medical record keeping, typically including outpatient medical records, inpatient progress notes, chief complaint text in structured forms, and present medical history text. This multi-source heterogeneous input data can be retrieved and read from the sleep center reporting system or electronic medical record attachment database via system interface integration. Using multi-page PSG reports with mixed text and images, combined with clinical text, maximizes the preservation of spatial and topological relationships in medical documents, avoiding information loss caused by traditional manual simplification and numerical extraction.

[0036] Step 102: Initialize the explicit evidence pointer library.

[0037] Step 103: Input the polysomnography report, the corresponding clinical document, and the explicit evidence pointer library into the sleep disorder auxiliary analysis model for parsing, and obtain the structured analysis results output by the sleep disorder auxiliary analysis model, which includes at least the analysis result field, evidence atom field, explanatory field, and adjustment suggestion field.

[0038] The evidence atom field contains at least one evidence atom to support the analysis result field, and the evidence pointer field of each evidence atom is taken from the evidence pointer in the explicit evidence pointer library. The sleep disorder auxiliary analysis model is trained on the report reading alignment model with the first input sample and its corresponding explicit evidence pointer library as input and the structured analysis label as the supervision signal. The report reading alignment model is trained with the polysomnography report sample and the sleep index label corresponding to the polysomnography report sample. The first input sample includes the polysomnography report sample and the clinical document corresponding to the polysomnography report sample.

[0039] In this context, the evidence pointer field of each evidence atom is taken from the evidence pointer in the explicit evidence pointer library. This can be understood as restricting the candidate evidence pointers to the legal evidence pointers in the explicit evidence pointer library when generating the evidence pointer field.

[0040] The sleep disorder auxiliary analysis model is a large model with multimodal image and text understanding capabilities. Internally, it performs a parsing process that fuses and understands the visual features of multi-page images with the text features of clinical documents. To ensure automated integration and verification with downstream hospital information systems, the parsed output is not freely generated long text, but is forcibly constrained into structured data containing specific fields.

[0041] Specifically, the Analysis Results field is used to represent the structured medical conclusions after comprehensive evaluation of the examinee; the Evidence Atom field is used to carry the set of atomic-level evidence after decomposition, ensuring that each piece of evidence is independent, concise and carries a single clinical fact; the Explanation field is mainly used to carry auxiliary explanations that cannot be directly used as numerical evidence, such as medical explanations of missing or contradictory data; and the Adjustment Suggestion field is used to output targeted clinical treatment suggestions or further follow-up examination plans.

[0042] In some embodiments, the adjustment suggestion field is used to output adjustment suggestions, treatment suggestions, follow-up suggestions, or further inspection suggestions.

[0043] In some embodiments, the suggested field is adjusted to free text or semi-structured text.

[0044] In some embodiments, the adjustment suggestion field can be further subdivided into subfields such as medication suggestions, behavioral suggestions, examination suggestions, and referral suggestions.

[0045] Building upon this foundation, the evidence atom field contains at least one evidence atom to support the analysis result field, and the evidence pointer field of each evidence atom is taken from the evidence pointer in the explicit evidence pointer library. This means that every piece of analytical evidence output by the model must belong to a valid pointer in the initialized explicit evidence pointer library, achieving fact-level traceability beyond the syntactic level.

[0046] The sleep disorder auxiliary analysis model employs a separate two-stage technical framework. It does not directly learn complex reasoning end-to-end from a general state. In the first stage, a report reading alignment model is pre-trained on polysomnography report samples with complex layouts and their corresponding objective sleep indicator labels. This model solves the problem of missed or misreading of cross-page tables and high-density trend charts, and develops a stable ability to extract objective physiological indicators. In the second stage, inheriting the capabilities of the report reading alignment model, clinical documents are introduced as the first joint input sample. Under the constraints of an explicit evidence pointer library, fine-tuning is performed using structured analysis labels annotated by doctors, ultimately generating the sleep disorder auxiliary analysis model.

[0047] In this embodiment, by introducing and initializing an explicit evidence pointer library, the atomic fields of evidence in the output structured analysis results must belong to this pointer library, thus limiting the degree of freedom of the model in generating evidence. Simultaneously, the two-stage parameter training of the sleep disorder auxiliary analysis model significantly improves the accuracy of its internal multimodal feature extraction when processing high-density, multi-page mixed text and image PSG report data, avoiding the omission and misinterpretation of basic factual data. Furthermore, it reduces the risk of unfounded claims propagating downstream and improves the traceability and auditability of the structured output.

[0048] In some embodiments, the evidence pointers in the explicit evidence pointer library include: Reporting indicator pointers are used to identify standardized indicators in polysomnography reports, and clinical text pointers are used to identify standardized anchor points in clinical texts.

[0049] Reporting indicator pointers are data indexes set for polysomnography reports, a visual or graphical modality of data. Since raw polysomnography reports typically contain complex graphs and tables, a set of standardized indicators is predefined. Standardized indicators are, at the system's underlying level, assigned a unified, machine-recognizable namespace or entity identifier to core objective physiological parameters in sleep medicine. Examples include the apnea-hypopnea index, minimum oxygen saturation, and the percentage of each sleep phase. Reporting indicator pointers uniquely identify and point to these standardized indicator entities during data processing. This allows the model to associate numerical values ​​extracted from the image features of a multi-page polysomnography report with the corresponding reporting indicator pointers, rather than freely generating vague textual descriptions such as "the patient's breathing is not good."

[0050] Meanwhile, clinical text pointers are data indexes set for the linguistic and textual modal data of the examinee's clinical documents. A normalized anchor point refers to a standardized medical history fragment, symptom tag, or key medical concept entity extracted by the system after preprocessing or entity recognition of unstructured clinical text. Examples include a history of snoring, daytime sleepiness, and a history of hypertension. Clinical text pointers are used to explicitly link the final analytical conclusions back to the corresponding normalized anchor point positions in the original medical record text during data flow within the system.

[0051] For example, when the explicit evidence pointer library is loaded, it may contain pre-defined indexes such as the PSG_Metric_AHI reporting index pointer and the Clinical_Symptom_Snoring clinical text pointer, which can be referenced by the sleep disorder auxiliary analysis model when generating structured analysis results.

[0052] In this embodiment, by constructing report indicator pointers and clinical text pointers in an explicit evidence pointer library, heterogeneous polysomnography image feature data and clinical history text feature data are uniformly mapped to the same standardized pointer namespace. When the sleep disorder auxiliary analysis model predicts and outputs structured analysis results from the input, it seeks physiological data evidence from the corresponding report indicator pointers and medical history data evidence from the corresponding clinical text pointers. This ensures that the final generated structured analysis results can associate reports and clinical texts, helping downstream systems to directly retrieve corresponding chart parameters and medical record paragraphs across modalities during automated data parsing, thereby improving the accuracy of joint inference using heterogeneous medical data.

[0053] In some embodiments, the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence further includes: Based on the normalization results of indicator names, indicator values, units, page positions, table fields, and chart areas in the polysomnography report, generate report indicator pointers; Anchor points are extracted from the chief complaint, present illness, symptoms, and medical history fragments in the clinical files corresponding to the polysomnography report to generate clinical text pointers. A library of explicit evidence pointers was constructed based on reporting indicator pointers and clinical text pointers.

[0054] In this embodiment, for the sleep scenario, an explicit evidence pointer library can be constructed based on polysomnography reports and corresponding clinical documents. This library can then be used as input to an auxiliary analysis model for sleep disorders to obtain structured analysis results.

[0055] In some embodiments, the evidence atom includes at least the following information: The Fact Claim field is used to describe the factual information claimed by the evidence atom; The source type field describes the source type of the evidence atom; The Evidence Pointer field describes the evidence pointer referenced by the evidence atom.

[0056] In this embodiment, when the sleep disorder auxiliary analysis model performs analysis, firstly, a single piece of extracted or derived clinical information is written into the fact claim field. This field describes the factual information claimed by the evidence atom. To avoid the problem of multiple difficult-to-verify statements mixed in long sentences due to the traditional free text mode, constraints are imposed on the fact claim field, requiring it to carry only a short, verifiable, single clinical analysis. For example, when the sleep disorder auxiliary analysis model extracts a polysomnography data item from the first input sample, it will directly assign the brief fact that the patient's apnea-hypopnea index is 35 breaths / hour to this fact claim field, without allowing other medical condition analyses to be mixed in.

[0057] During the extraction or derivation process, the source of the claimed factual information is determined, and the corresponding classification identifier is written into the source type field. This field describes the source type of the evidence atom. The data format of the source type field is typically limited to a predefined classification enumeration value, such as being strictly divided into clinical text, report indicators, derivation results, or missing information. The use of the source type field allows external systems to immediately identify the source of the factual claim when reading the object.

[0058] After determining the source type field, the valid dictionary key or addressing identifier strongly associated with the fact claim is written into the evidence pointer field. This field describes the evidence pointer referenced by the evidence atom. Simultaneously with outputting the aforementioned fact claim, the corresponding normalized identifier code must be retrieved from the initialized explicit evidence pointer library and configured in the evidence pointer field. The use of the evidence pointer field explicitly expresses the evidence pointer referenced by the fact claim field.

[0059] In this embodiment, the evidence atom is represented by a structured representation of the fact claim field, the source type field, and the evidence pointer field. While accurately expressing the evidence atom fields, it can improve the parsability and field stability of the structured analysis results, making it easier for downstream systems to consume directly and reducing interface incompatibility issues caused by fluctuations in free text.

[0060] In some embodiments, the description field includes at least the following information: The description item's type information, encoding information, target information, evidence pointer field referenced by the description item, and description text. The type information of the description item is one of the following types: Missing information, supporting information, conflicting information, next steps suggestions, and conservative output information; Among them, the conservative output information is used to indicate that due to insufficient evidence, it is impossible to make a judgment and to recommend manual review.

[0061] For the type information of the description item, when constructing the field, its value will be forcibly mapped to a preset small closed category set. This category set is specifically limited to one of five types: missing information, supporting information, conflicting information, next step suggestion, and conservative output information.

[0062] Specifically, if the polysomnography report detects incomplete key respiratory event channel data or lacks necessary medical history anchors in the clinical file, the sleep disorder auxiliary analysis model will output a description item with missing information type; if the polysomnography report indicators are found to be logically inconsistent with the chief complaint in the clinical file, a description item with conflict information type will be output; similarly, the model can also generate supporting information based on existing evidence or output next step suggestions for supplementary examinations.

[0063] The coded information for the description item refers to the unique identifier or medical standard mapping code assigned to that description item to facilitate rapid retrieval and machine parsing by downstream hospital management information systems or electronic medical record systems. Its scope can encompass system-defined internal status codes or standard International Classification of Disease codes extensions.

[0064] The target information of the explanatory item is used to establish a topological relationship between the explanatory item and the specific conclusion. That is, this information field explicitly indicates which specific dimension the current explanatory item is used to explain. For example, if the system provides the analysis result for obstructive sleep apnea, the target information of the explanatory item will point to that specific analysis result field, making the data tightly linked at the logical level.

[0065] Although the explanatory text carries the function of auxiliary interpretation, the evidence pointer field referenced by the explanatory text still requires that the underlying data source that triggers the explanatory text be recorded in the form of a pointer, and that the pointer must come from the aforementioned initialized explicit evidence pointer library.

[0066] The explanatory text for the explanatory items is a piece of free text that doctors can read, output by the large language model branch within the sleep disorder auxiliary analysis model based on the above structured state and using its natural language generation capabilities. It is used to present specific explanations to doctors on the clinical interface.

[0067] For example, when a polysomnography report contains missing oxygen saturation data due to sensor detachment, the sleep disorder auxiliary analysis model will not forcibly fabricate an oxygen saturation value for analysis during inference. Instead, the model will generate a description, where the type information is missing information, the target information points to the respiratory event dimension, the cited evidence pointer field points to the anomaly pointer corresponding to the oxygen saturation channel in the report, and the description text is generated in natural language to explain that the severity of hypoxemia cannot be assessed due to sensor detachment, thus triggering the system's conservative output strategy.

[0068] In this embodiment, by setting a data structure for type information, encoding information, target information, evidence pointer fields, and explanatory text in the description field, this mandatory data structure setting can prevent forced inferences to complete the answer when the sleep disorder auxiliary analysis model encounters incomplete or conflicting multi-source heterogeneous data. It triggers conservative strategies by filling in missing information. Simultaneously, the introduction of target information and cited evidence pointers ensures that the generated explanatory text not only has good clinical readability but also forms a binding relationship with the pointer library at the underlying data structure level, improving the security and verifiability of the data output.

[0069] In some embodiments, the number of evidence atoms contained in the evidence atom field is less than or equal to a preset number threshold.

[0070] In this embodiment, a preset quantity threshold is used as the upper limit of the number of evidence atoms. This threshold can be dynamically changed. The use of the upper limit of the number of evidence atoms prevents the model from piling up too much loose evidence and prompts it to prioritize the most critical and discriminative evidence items, thereby improving verifiability.

[0071] In some embodiments, the preset quantity threshold is set to 6.

[0072] In some embodiments, the upper limit of the number of evidence atoms can be dynamically allocated according to the analysis dimension, adaptively set according to the disease type, or adjusted according to the page complexity; deduplication constraints, minimum coverage constraints, and priority sorting constraints can also be added.

[0073] like Figure 2 As shown, in some embodiments, the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence further includes: Step 201: Verify the structured analysis results using preset verification items.

[0074] Step 202: If the structured analysis results pass the verification, output the structured analysis results.

[0075] The preset verification items include one or more of the following: The system performs the following checks: integrity of structured format; whether the evidence atom field contains evidence atoms; whether the evidence pointer belongs to the pointer attribution of the explicit evidence pointer library; number of evidence atoms corresponding to the evidence atoms contained in the evidence atom field; consistency of structured analysis results; missing information check; whether to output conservative structured analysis results when the evidence atom field does not contain evidence atoms; and whether there is a conflict between missing information in the description field and the analysis result field.

[0076] Among them, structured format integrity verification: This verification method refers to using a preset schema parser (such as a JSONSchema validator) to parse the output structured analysis results, and determine whether they meet the predetermined key-value pair structure, whether there are missing fields, and whether the hierarchy is consistent, so as to ensure that the output data has good machine-readable parsing.

[0077] In some embodiments, the output structured analysis results may also be in the format of Extensible Markup Language (XML), YAML, protocol buffer, form object, database record structure, or other machine-readable format, wherein YAML is a lightweight, data-centric serialization language.

[0078] Pointer attribution verification to determine if an evidence pointer belongs to the explicit evidence pointer library: This verification method involves the system extracting all reference pointers contained in the structured analysis results and comparing them with the set of legitimate pointers loaded during the initialization phase. If it is found that the model has fabricated a phantom pointer that is not in the explicit evidence pointer library, the verification is deemed to have failed, thereby achieving a strong binding of legitimate sources.

[0079] Verification of the number of evidence atoms corresponding to the evidence atoms contained in the evidence atom field: This verification method refers to the system counting the evidence entries within the output evidence atom field. By comparing the statistical count with the preset system budget limit, it prevents the model from piling up redundant, loose, and invalid evidence to meet the requirements of completeness, thus ensuring the compactness of the evidence set.

[0080] Structured analysis result content consistency verification: This verification method refers to the system performing cross-logic verification between various structured fields to determine whether there are isolated conclusions in the analysis result fields that lack atomic binding with evidence, or whether there are semantic or semantic contradictions between the supporting information and the analysis conclusions.

[0081] Information Missing Validation: This validation method is a built-in security rejection mechanism of the system. Specifically, based on the input multi-source data and the output description, it judges whether there are abnormal situations such as missing or insufficient evidence of key respiratory events, but the model still forcibly gives a positive analysis conclusion, as the basis for triggering the conservative output strategy.

[0082] The structured analysis result will only be released and output to the downstream hospital management information system or electronic medical record system if all the pre-defined verification items return a valid security signal. If any verification fails, an interception mechanism will be automatically triggered, automatically rejecting the forced analysis or placing it in the manual review queue.

[0083] Specifically, if the structured analysis result fails the verification, the output of the structured analysis result is canceled, and a conservative structured analysis result, information on missing items in the conservative structured analysis result, and result prompt information are generated. The result notifications include prompts for insufficient evidence or prompts for manual review.

[0084] The conservative structured analysis result is the result output by triggering the conservative output strategy. It only contains the partial structured analysis results that can be given. For missing structured analysis results, the missing items in the conservative structured analysis result are explained, and result prompt information is given.

[0085] In this embodiment, the use of preset verification items reduces the risk of unfounded claims spreading downstream and improves the traceability and auditability of structured outputs, achieving conservative handling when key information is uncertain.

[0086] In some embodiments, such as Figure 3 and Figure 5 As shown, the report reading alignment model is trained through the following steps: Step 301: Input the polysomnography report samples and the pre-built structured extraction constraint prompts into the initial model to extract indicators and obtain the extraction results corresponding to each indicator.

[0087] The initial model typically refers to a multimodal foundational model that has not undergone alignment training in the first stage. Pre-built structured extraction constraint prompts are strictly defined system instruction texts that only require the model to extract objective physiological indicators and their corresponding page coordinates or chart positions, deliberately shielding any inference-related guidance instructions. These structured extraction constraints are used to constrain the content contained in the extraction results output by the initial model.

[0088] Specifically, the multichannel sleep map report samples, which are visual modalities, and the cue words, which are text modalities, are simultaneously input into the initial model. The initial model decodes based on the current strategy and outputs structured basic data objects as extraction results.

[0089] Step 302: Using the sleep index labels corresponding to the polysomnography report samples as a reference, a rule-based scorer is used to evaluate each extraction result and obtain a score value for each extraction result.

[0090] In this step, the rule-based scorer is an automated evaluation mechanism that introduces hard-sounding logical rules, rather than a black-box evaluation model relying on external value networks. The generated extraction results are automatically and accurately compared with the doctor's pre-labeled first-stage sleep indicator tags, i.e., real objective data tags. The evaluation process can be broken down into several specific items, from broad to narrow: for example, a format compliance score to check whether output fields are missing and whether the data type conforms to the preset structured pattern; a numerical accuracy score to calculate the precise error between the indicator values ​​extracted by the model and the label values; and a cross-page consistency score to check whether there are logical conflicts between related indicators. Through these multi-dimensional hard-sounding rule calculations, the rule-based scorer assigns a clear absolute score value to each extraction result within each group.

[0091] Step 303: Based on the scores of all the extraction results, optimize the initial model using a pre-selected model optimization algorithm until the optimization conditions are met.

[0092] In this step, the model optimization algorithm can be one of the following: group relative policy optimization algorithm, rule-driven reinforcement learning, preference optimization, rejection sampling distillation, reward model distillation, or supervised + rule hybrid training.

[0093] In some embodiments, taking a group-based relative policy optimization algorithm as an example, standardized comparisons are performed within the N candidate extraction results generated from the same input sample: first, the mean and standard deviation of all scores within the current group are calculated; then, a relative advantage value is calculated for each extraction result. When the relative advantage value of an extraction is positive, a truncation mechanism is used to encourage an increase in the generation probability of that output; when it is negative, a suppression penalty is applied. Subsequently, the gradient of the loss function is calculated based on the relative advantage value, and the weight parameters of the initial model are updated through backpropagation.

[0094] During this parameter update process, to prevent the model from catastrophically forgetting information in pursuit of high scores based on rules, thus damaging its underlying language capabilities, a reference model with frozen parameters is introduced. The KL divergence between the current strategy and the output probability distribution of the reference model is dynamically calculated as a penalty term for regularization constraints. After multiple iterations, when the model's loss function converges or reaches preset optimization conditions such as the maximum number of training steps, the parameters are fixed, resulting in a report reading alignment model with highly stable chart reading capabilities.

[0095] In this embodiment, by introducing restricted prompt words that are only extracted for objective physiological indicators during the model training phase, combined with the hard comparison calculation mechanism of the rule scorer in multiple dimensions such as format, numerical value, and cross-page consistency, and by adopting a model optimization algorithm based on the extraction result score value in the parameter update layer, the problem of visual attention drift when multimodal large models are processed in multi-page, dense data tables is solved, thereby improving the stability and consistency of indicator extraction of the sleep disorder auxiliary analysis model.

[0096] In one specific implementation, for the same input sample, the initial model generates a set of candidate extraction results containing multiple groups of relative samples in parallel to form a set of comparative data for subsequent optimization.

[0097] In some embodiments, the explicit evidence pointer library may also include hierarchical or grouped levels such as page pointers, table pointers, chart pointers, chapter pointers, medical history pointers, and medication pointers; it may also be represented by hash pointers, enumeration identifiers, or tree-like namespaces.

[0098] In some embodiments, such as Figure 4 and Figure 5 As shown, the sleep disorder auxiliary analysis model is trained through the following steps: Step 401: Input the polysomnography report sample from the first input sample into the visual encoder for feature extraction to obtain a visual feature map sequence.

[0099] In this step, multi-page polysomnography report samples are input into a visual encoder for feature extraction. The visual encoder extracts features from the index tables, sleep structure trend graphs, and waveforms spanning multiple pages in the report using low-level convolutional kernels or self-attention mechanisms, thereby outputting a sequence of visual feature maps containing high-dimensional feature information. This sequence fully preserves the spatial topological relationship of the original charts on the two-dimensional physical page in terms of data structure.

[0100] Step 402: Use the visual-language fusion module to convert the visual feature map sequence into a visual word sequence.

[0101] Since the high-dimensional representation space of the visual feature map sequence is not interconnected with the text language space, a visual-language fusion module is needed to reduce the dimensionality of the extracted high-dimensional visual feature map sequence and perform semantic space mapping. After linear or nonlinear transformation processing by this module, the visual feature map is transformed into a visual word sequence that is completely consistent with the text embedding feature dimension of the large language model, thereby achieving word alignment within a unified representation space.

[0102] Step 403: Input the clinical documents and explicit evidence pointer library corresponding to the polysomnography report samples into the large language model for mapping to obtain the text lexical sequence.

[0103] In this step, clinical documents containing the chief complaint text, present medical history text, and an explicit evidence pointer library are uniformly input into the embedding layer of the large language model. The embedding layer of the large language model maps the aforementioned text symbols and predefined legal pointer sets into a high-dimensional vector set that the model can compute, thereby obtaining a text lexical sequence, through table lookup and positional encoding operations.

[0104] Step 404: After concatenating the visual lexical sequence and the text lexical sequence, input them into the large language model for prediction to obtain structured analysis prediction.

[0105] The aligned visual lexical sequence and the textual lexical sequence are concatenated along the time step dimension to form a unified joint multimodal context sequence. The large language model uses this joint sequence as historical input and employs an autoregressive decoding method to predict the output sequence. In specific training states, a teacher-mandated strategy is typically adopted, where, when the model predicts each lexical step, real structured analysis labels manually annotated by a doctor and strictly selected from a pointer library are used as historical context prompts to guide the large language model to propagate forward step by step, ultimately outputting a structured analysis prediction sequence.

[0106] Step 405: Update at least one of the first trainable parameter, the second trainable parameter, and the third trainable parameter based on the cross-entropy loss of the structured analysis prediction and the structured analysis label corresponding to the first input sample, until the training conditions are met.

[0107] In some embodiments, the low-rank adaptation parameter in the second trainable parameters is updated based on the cross-entropy loss of the structured analysis prediction and the structured analysis label corresponding to the first input sample.

[0108] The first trainable parameter is the trainable parameter of the visual encoder, the second trainable parameter is the trainable parameter of the large language model, and the third trainable parameter is the trainable parameter of the vision-language fusion module.

[0109] In this step, the cross-entropy loss between the structured analysis predictions output by the large language model and the true structured analysis labels is calculated. During training or decoding, through legal pointer label supervision, candidate pointer masking, constrained decoding, illegal pointer penalty terms, or posterior pointer attribution verification, the evidence pointers output by the model are restricted to legal pointers in the explicit evidence pointer library. This cross-entropy loss is backpropagated using optimizers such as gradient descent. Alternatively, efficient parameter fine-tuning can be employed, by freezing most of the weights of the first trainable parameter and updating the gradients only for the third and second trainable parameters, until the overall loss converges and the preset training termination condition is met.

[0110] In this embodiment, a sleep disorder auxiliary analysis model can be trained to output structured analysis results including analysis result fields, evidence atom fields, explanatory field fields, and adjustment suggestion fields.

[0111] In some embodiments, the sleep disorder auxiliary analysis model can employ supervised fine-tuning, parameter-efficient fine-tuning, or full-parameter fine-tuning to force teacher learning on the structured target object resulting from the concatenation of visual word sequences and text word sequences.

[0112] In some embodiments, the multimodal model consisting of a visual encoder, a visual-language fusion module, and a large language model can also be replaced by a combination of a self-developed visual encoder and a language model for the medical field.

[0113] The following describes the traceable sleep disorder auxiliary analysis device based on pointer-anchored evidence provided by the present invention. The traceable sleep disorder auxiliary analysis device based on pointer-anchored evidence described below can be referred to in correspondence with the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence described above.

[0114] This invention proposes a traceable sleep disorder auxiliary analysis device based on pointer-anchored evidence, such as... Figure 6 As shown, it includes the following modules: The acquisition module 601 is used to acquire the polysomnography report of the target subject and the corresponding clinical documents of the polysomnography report; Initialization module 602 is used to initialize the explicit evidence pointer library; The parsing module 603 is used to input the polysomnography report, the corresponding clinical documents and explicit evidence pointer library into the sleep disorder auxiliary analysis model for parsing, and to obtain the structured analysis results output by the sleep disorder auxiliary analysis model, which includes at least the analysis result field, evidence atom field, explanatory field and adjustment suggestion field. The evidence atom field contains at least one evidence atom to support the analysis result field, and the evidence pointer field of each evidence atom is taken from the evidence pointer in the explicit evidence pointer library. The sleep disorder auxiliary analysis model is trained on the report reading alignment model with the first input sample and its corresponding explicit evidence pointer library as input and the structured analysis label as the supervision signal. The report reading alignment model is trained with the polysomnography report sample and the sleep index label corresponding to the polysomnography report sample. The first input sample includes the polysomnography report sample and the clinical document corresponding to the polysomnography report sample.

[0115] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740. The processor 710 can call logic instructions in the memory 730 to execute a traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence. This method includes: acquiring the polysomnography report of the target subject and the corresponding clinical document; initializing an explicit evidence pointer library; inputting the polysomnography report, the corresponding clinical document, and the explicit evidence pointer library into a sleep disorder auxiliary analysis model for parsing, and obtaining a structured analysis result output by the sleep disorder auxiliary analysis model, which includes at least an analysis result field, an evidence atom field, a description field, and an adjustment suggestion field. The evidence atom field contains at least one evidence atom to support the analysis result field, and the evidence pointer field of each evidence atom is taken from the evidence pointer in the explicit evidence pointer library. The sleep disorder auxiliary analysis model is trained using a first input sample and its corresponding explicit evidence pointer library as input, and a structured analysis label as a supervision signal to train a report reading alignment model. The report reading alignment model is trained using a polysomnography report sample and the sleep index label corresponding to the polysomnography report sample. The first input sample includes the polysomnography report sample and the corresponding clinical document.

[0116] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence provided by the above methods. The method includes: acquiring the polysomnography report of the target subject and the corresponding clinical documents; initializing the explicit evidence pointer library; inputting the polysomnography report, the corresponding clinical documents, and the explicit evidence pointer library into the sleep disorder auxiliary analysis model for parsing, and obtaining the output of the sleep disorder auxiliary analysis model, which includes at least the analysis result fields. The structured analysis results include evidence atom fields, explanatory fields, and adjustment suggestion fields. Each evidence atom field contains at least one evidence atom to support the analysis result fields, and the evidence pointer field for each evidence atom is taken from the explicit evidence pointer library. The sleep disorder auxiliary analysis model is trained on the report reading alignment model using the first input sample and its corresponding explicit evidence pointer library as input, and the structured analysis labels as supervision signals. The report reading alignment model is trained using polysomnography report samples and the sleep index labels corresponding to the polysomnography report samples. The first input sample includes polysomnography report samples and the corresponding clinical documents.

[0118] Furthermore, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program performs the aforementioned method for traceable sleep disorder auxiliary analysis based on pointer-anchored evidence. This method includes: acquiring a polysomnography report of a target subject and the corresponding clinical document; initializing an explicit evidence pointer library; and inputting the polysomnography report, the corresponding clinical document, and the explicit evidence pointer library into a sleep disorder auxiliary analysis model for parsing, thereby obtaining an output from the sleep disorder auxiliary analysis model that includes at least an analysis result field, an evidence atom field, and a description field. The structured analysis results of the adjustment suggestion field; wherein, the evidence atom field contains at least one evidence atom to support the analysis result field, and the evidence pointer field of each evidence atom is taken from the evidence pointer in the explicit evidence pointer library. The sleep disorder auxiliary analysis model is trained on the report reading alignment model with the first input sample and its corresponding explicit evidence pointer library as input and the structured analysis label as the supervision signal. The report reading alignment model is trained with the polysomnography report sample and the sleep index label corresponding to the polysomnography report sample. The first input sample includes the polysomnography report sample and the clinical document corresponding to the polysomnography report sample.

[0119] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0120] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence, characterized in that, include: Obtain the polysomnography report of the target subject and the corresponding clinical documents of the polysomnography report; Initialize the explicit evidence pointer library; The polysomnography report, the corresponding clinical document, and the explicit evidence pointer library are input into the sleep disorder auxiliary analysis model for parsing, and the structured analysis results output by the sleep disorder auxiliary analysis model include at least the analysis result field, evidence atom field, explanatory field, and adjustment suggestion field. The evidence atom field contains at least one evidence atom to support the analysis result field, and the evidence pointer field of each evidence atom is taken from the evidence pointer in the explicit evidence pointer library. The sleep disorder auxiliary analysis model is trained on the report reading alignment model using the first input sample and its corresponding explicit evidence pointer library as input and the structured analysis label as the supervision signal. The report reading alignment model is trained on the polysomnography report sample and the sleep index label corresponding to the polysomnography report sample. The first input sample includes the polysomnography report sample and the clinical document corresponding to the polysomnography report sample.

2. The method for traceable sleep disorder auxiliary analysis based on pointer-anchored evidence according to claim 1, characterized in that, The evidence pointers in the explicit evidence pointer library include: Reporting indicator pointers are used to identify normalized indicators in the polysomnography report, and clinical text pointers are used to identify normalized anchor points in the clinical text.

3. The method for traceable sleep disorder auxiliary analysis based on pointer-anchored evidence according to claim 2, characterized in that, The traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence also includes: Based on the normalization results of the indicator names, indicator values, units, page positions, table fields, and chart areas in the polysomnography report, the report indicator pointers are generated. Anchor points are extracted from the chief complaint, present illness, symptoms, and medical history fragments in the clinical file corresponding to the polysomnography report to generate the clinical text pointer; The explicit evidence pointer library is constructed based on the reported indicator pointers and the clinical text pointers.

4. The method for traceable sleep disorder auxiliary analysis based on pointer-anchored evidence according to claim 1, characterized in that, The evidence atom includes at least the following information: The Fact Claim field is used to describe the factual information claimed by the evidence atom; The source type field describes the source type of the evidence atom; The Evidence Pointer field describes the evidence pointer referenced by the evidence atom.

5. The method for traceable sleep disorder auxiliary analysis based on pointer-anchored evidence according to claim 1, characterized in that, The description field must include at least the following information: The description item's type information, encoding information, target information, evidence pointer field referenced by the description item, and description text. The type information of the description item is one of the following types: Missing information, supporting information, conflicting information, next steps suggestions, and conservative output information; The conservative output information is used to indicate that due to insufficient evidence, a judgment cannot be made and manual review is recommended.

6. The method for traceable sleep disorder auxiliary analysis based on pointer-anchored evidence according to claim 1, characterized in that, The number of evidence atoms contained in the evidence atom field is less than or equal to a preset number threshold.

7. The method for traceable sleep disorder auxiliary analysis based on pointer-anchored evidence according to claim 1, characterized in that, The traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence also includes: The structured analysis results are verified using preset verification items; If the structured analysis results pass the verification, the structured analysis results are output. If the structured analysis result fails the verification, the output of the structured analysis result is canceled, and a conservative structured analysis result, information on missing items in the conservative structured analysis result, and result prompt information are generated. The result prompts include prompts for insufficient evidence or prompts for manual review; The preset verification items include one or more of the following: The following checks are performed: structured format integrity verification; whether the evidence atom field contains evidence atoms; whether the evidence pointer belongs to the pointer attribution of the explicit evidence pointer library; the number of evidence atoms corresponding to the evidence atoms contained in the evidence atom field; consistency verification of the structured analysis result content; information missing verification; whether to output conservative structured analysis results when the evidence atom field does not contain evidence atoms; and whether there is a conflict between the missing information in the description field and the analysis result field.

8. The method for traceable sleep disorder auxiliary analysis based on pointer-anchored evidence according to any one of claims 1 to 7, characterized in that, The report reading alignment model was trained through the following steps: The polysomnography report samples and pre-built structured extraction constraint prompts are input into the initial model to extract indicators, and the extraction results corresponding to each indicator are obtained. Using the sleep index labels corresponding to the polysomnography report samples as a reference, a rule-based scorer is used to evaluate each extraction result to obtain a score value for each extraction result. Based on the scores of all the extraction results, the initial model is optimized using a pre-selected model optimization algorithm until the optimization conditions are met; The structured extraction constraints are used to constrain the content contained in the extraction results output by the initial model.

9. The method for traceable sleep disorder auxiliary analysis based on pointer-anchored evidence according to any one of claims 1 to 7, characterized in that, The sleep disorder auxiliary analysis model was trained through the following steps: The polysomnography report sample in the first input sample is input into the visual encoder for feature extraction to obtain a visual feature map sequence. The visual feature map sequence is converted into a visual word sequence using a visual-language fusion module; The clinical documents corresponding to the polysomnography report samples and the explicit evidence pointer library are input into a large language model for mapping to obtain a text word sequence; The visual word sequence and the text word sequence are concatenated and input into the large language model for prediction to obtain structured analysis prediction; Based on the cross-entropy loss of the structured analysis prediction and the structured analysis label corresponding to the first input sample, update at least one of the first trainable parameter, the second trainable parameter, and the third trainable parameter until the training conditions are met. The first trainable parameter is the trainable parameter of the visual encoder, the second trainable parameter is the trainable parameter of the large language model, and the third trainable parameter is the trainable parameter of the visual-language fusion module.

10. A traceable sleep disorder auxiliary analysis device based on pointer-anchored evidence, characterized in that, include: The acquisition module is used to acquire the polysomnography report of the target subject and the corresponding clinical documents of the polysomnography report; The initialization module is used to initialize the explicit evidence pointer library; The parsing module is used to input the polysomnography report, the corresponding clinical document of the polysomnography report and the explicit evidence pointer library into the sleep disorder auxiliary analysis model for parsing, and to obtain the structured analysis results output by the sleep disorder auxiliary analysis model, which includes at least the analysis result field, evidence atom field, explanatory field and adjustment suggestion field. The evidence atom field contains at least one evidence atom to support the analysis result field, and the evidence pointer field of each evidence atom is taken from the evidence pointer in the explicit evidence pointer library. The sleep disorder auxiliary analysis model is trained on the report reading alignment model using the first input sample and its corresponding explicit evidence pointer library as input and the structured analysis label as the supervision signal. The report reading alignment model is trained on the polysomnography report sample and the sleep index label corresponding to the polysomnography report sample. The first input sample includes the polysomnography report sample and the clinical document corresponding to the polysomnography report sample.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the traceable sleep disorder auxiliary analysis method based on pointer-anchored evidence as described in any one of claims 1 to 9.

Images