System for early screening and diagnosis of cognitive impairment based on multi-modal data

Abstract

The application discloses a kind of cognitive impairment early screening and diagnosis system based on multi-modal data, including following module: (1) multi-modal data acquisition module, for obtaining image data, gene data, behavior data and clinical narrative information;The clinical narrative information includes at least two kinds of patient complaint text, interference factor hook item, symptom time axis;(2) causal purification module, for filtering the pseudo-correlation between multi-modal data by clinical narrative anchoring and reverse intervention deduction, including: narrative anchoring unit: based on preset causal conflict rule base, through structured clinical narrative field identification data conflict, trigger freezing or weight reduction operation;Intervention deduction unit: generate the reverse intervention verification package containing low-cost intervention measures and review plan, and update causal rule base based on review data;The application breaks through the statistical correlation limitation of traditional technology, realizes the upgrade of diagnosis and treatment paradigm from "correlation" to "causality".

Description

Technical Field

[0001] This invention belongs to the field of medical diagnostics, specifically a system for early screening and diagnosis of cognitive impairment based on multimodal data. Background Technology

[0002] In the field of early screening and diagnosis of cognitive impairment, multimodal data fusion technology has become a research hotspot. However, existing technologies face significant challenges in handling the complex correlations of multimodal data, mainly manifested in the following limitations:

[0003] 1. Traditional methods rely on statistical correlations to establish multimodal associations, but they cannot distinguish between true causality and spurious correlations. For example, APOE4 gene mutations may indirectly lead to cognitive decline by accelerating brain atrophy, but models often directly attribute brain atrophy to cognitive impairment, ignoring the mediating role of genes. Similarly, gait abnormalities may be caused by orthopedic diseases such as arthritis, but are misdiagnosed as a marker of cognitive decline. Such misdiagnosis leads to models relying on spurious features, resulting in a high misdiagnosis rate in clinical applications and a significant decrease in model generalization ability.

[0004] 2. Existing systems mostly rely on expert experience to build fixed rule bases, but they cannot adapt to the following dynamic scenarios:

[0005] Drug interference: The use of benzodiazepines can mask the true genetic risks, and static rules make it difficult to adjust weights in a timely manner.

[0006] Individual differences: The same symptoms may be caused by different etiologies (e.g., hippocampal atrophy can be caused by AD pathology or depression), and static models cannot be dynamically decoupled.

[0007] Such rigid design leads to a disconnect between the rule base and clinical practice, requiring frequent manual corrections and resulting in low efficiency.

[0008] 3. The core defects of existing technologies include:

[0009] Insufficient time perception: Ignoring the temporal sequence of pathological progression (e.g., brain atrophy precedes cognitive score decline) and mistakenly treating co-occurrence associations as causation.

[0010] Black box decision-making: Deep learning models lack interpretable mechanisms, making it impossible for doctors to verify the biological rationale behind feature associations, thus hindering clinical adoption.

[0011] 4. Traditional biomarker frameworks rely on large sample training data, but for rare subtypes such as posterior cortical atrophy AD, small samples (<50 cases) result in classification accuracy of less than 60%, making it difficult to support precision treatment. Summary of the Invention

[0012] The purpose of this invention is to provide an early screening and diagnosis system for cognitive impairment based on multimodal data. This invention breaks through the limitations of statistical correlation in traditional technologies and achieves an upgrade in the diagnostic and treatment paradigm from "correlation" to "causation".

[0013] The technical solution adopted in this invention is as follows:

[0014] An early screening and diagnostic system for cognitive impairment based on multimodal data includes the following modules:

[0015] (1) A multimodal data acquisition module, used to acquire image data, gene data, behavioral data and clinical narrative information; the clinical narrative information includes at least two of the following: patient complaint text, interfering factor checklist, and symptom timeline;

[0016] (2) Causal purification module, used to filter spurious correlations between multimodal data through clinical narrative anchoring and reverse intervention deduction, including:

[0017] Narrative anchoring unit: Based on a pre-defined causal conflict rule base, it identifies data conflicts through structured clinical narrative fields and triggers freeze or de-weighting operations;

[0018] Intervention simulation unit: Generates a reverse intervention validation package containing low-cost intervention measures and a review plan, and updates the causal rule base based on the review data;

[0019] (3) Dynamic fusion module, used for time-aware alignment of time-series heterogeneous data, and modeling the interaction effect between pathological decline and behavioral compensation, including:

[0020] Time-varying coding unit: captures the nonlinear relationship between brain atrophy rate and cognitive score in longitudinal multimodal data;

[0021] Explainable attention mechanisms: the weighting of pathological features and compensatory behaviors;

[0022] (4) Subtype classification module: a classification decision tree is constructed based on a biomarker framework, and a prototype network is used to handle the small sample learning problem of rare subtypes;

[0023] (5) Clinical interaction module, outputting two-color label reports, causal quick lookup tools and standardized intervention verification suggestion forms.

[0024] The causal conflict rule base is constructed through a keyword mapping relationship table, whereby the keywords include disease names, drug names, and keywords related to life events. When a keyword conflict is detected, the associated modal data is automatically frozen.

[0025] The reverse intervention verification package satisfies the following conditions:

[0026] The intervention does not conflict with the patient's current treatment plan;

[0027] The intervention cost is lower than the preset threshold;

[0028] The intervention effect is quantitatively evaluated using a formula:

[0029] Causality strength = (Δintervention group - Δcontrol group) / σbaseline

[0030] Wherein, Δintervention group is the rate of change of the target modality index in the intervention group, Δcontrol group is the rate of change of the matched control group, and σbaseline is the baseline standard deviation; causal strength > 0.5 is marked as strong causality, 0.2-0.5 is marked as weak causality, and < 0.2 is marked as spurious correlation.

[0031] The dual-color label report includes:

[0032] Red label: Data frozen due to keyword conflicts;

[0033] Yellow label: Marks associations that need to be verified.

[0034] The cause-and-effect quick lookup tool displays high-frequency conflict scenarios and handling rules in the form of interactive posters.

[0035] An electronic medical record system integrates the aforementioned early screening and diagnosis system for cognitive impairment based on multimodal data, and interacts with the hospital's HIS system in the following ways:

[0036] Automatically extract medication records and comorbidity information from the HIS as clinical narrative fields;

[0037] Embed the two-color label report into the examination results page of the electronic medical record.

[0038] A computer-readable storage medium storing program instructions that, when executed by a processor, enable the function of the early screening and diagnosis system for cognitive impairment based on multimodal data.

[0039] A clinical toolkit for screening cognitive impairment includes:

[0040] A tear-resistant, cause-and-effect quick reference poster, printed with high-frequency conflict rules and with a reserved area for handwritten addition of new rules;

[0041] Red and yellow bicolor labels are used for manually marking conflicting data in inspection reports;

[0042] Standardized intervention verification cards, including check-in intervention options and QR codes linking to the follow-up plan.

[0043] A cognitive impairment screening method based on multimodal data includes the following steps:

[0044] (1) Collect multimodal data and clinical narrative information, and trigger data freezing through keyword mapping;

[0045] (2) Dynamically fuse the remaining data and model the temporal decay and compensatory effects;

[0046] (3) Output subtype classification results and two-color label report, and recommend intervention verification measures;

[0047] (4) Update the causal rule base and classification model based on the review data.

[0048] The standardized intervention verification recommendation form includes checkbox options and a review plan template.

[0049] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0050] This invention systematically solves the problem of causal confusion caused by apparent correlation in multimodal data through a multi-layered technical architecture, specifically manifested in the following core mechanisms:

[0051] 1. This invention constructs a structured causal rule base, maps multimodal data relationships through preset keywords, and automatically freezes or reduces the weight of interfering factors. It matches clinical narrative information in real time, triggers rule execution, and generates traceable audit logs to block the impact of spurious associations on the model.

[0052] 2. This invention generates low-cost intervention measures, quantifies causal effects through controlled experiments, dynamically updates the rule base, and distinguishes between strong causality and spurious correlation. The model is continuously optimized based on review data to ensure that causal relationships evolve dynamically with new evidence.

[0053] 3. This invention utilizes LSTM networks to capture the longitudinal correlation between brain atrophy and cognitive scores, identifying temporal causal relationships. It distinguishes between pathological features and compensatory behaviors through attention mechanisms, avoiding interference from non-cognitive factors.

[0054] 4. This invention constructs a classification tree based on the ATN standard to distinguish typical disease subtypes. It uses a prototype network to process rare cases, solving the classification challenge under conditions of data scarcity.

[0055] 5. In summary, this invention upgrades correlation analysis to causal inference, reducing misdiagnosis and generalization failure. Through a rule base and model that continuously evolve with new evidence, it adapts to iterative developments in medical cognition. Visualization mechanisms and structured reports support clinical decision-making and promote doctor-patient collaboration. Its modular design can be transferred to other disease areas, driving interdisciplinary innovation in diagnosis and treatment. This system, through a full-link design of "constraint-validation-decoupling-interpretation," guides multimodal data from statistical correlation to biological causality, improving the closed-loop accuracy of cognitive impairment diagnosis and treatment from data to decision. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0058] See Figure 1 This invention relates to an early screening and diagnosis system for cognitive impairment based on multimodal data, comprising the following modules:

[0059] (1) A multimodal data acquisition module, used to acquire image data, gene data, behavioral data and clinical narrative information; the clinical narrative information includes at least two of the following: patient complaint text, interfering factor checklist, and symptom timeline;

[0060] Specifically, the information collected by the multimodal data acquisition module includes:

[0061] Imaging data: Brain structural images were acquired using a 3T MRI device. Hippocampal volume and cortical thickness were measured using a T1-MPRAGE sequence (TR=2300ms, TE=2.98ms, voxel 1×1×1mm³). The rate of brain atrophy was compared longitudinally (annual decline rate >5% triggers an early warning).

[0062] Genetic data: APOE4, PSEN1 and other gene loci were detected based on the sequencing platform. The bioinformatics analysis process included FastQC quality control, BWA alignment and ANNOVAR annotation. The results were integrated into the hospital HIS system through the HL7 protocol.

[0063] Behavioral data: Gait symmetry (triaxial accelerometer FFT analysis) and sleep quality (PPG heart rate variability monitoring) are collected through wearable devices, and the data is encrypted and uploaded to the hospital's private cloud every hour.

[0064] Clinical narrative information specifically includes:

[0065] Chief complaint text: The doctor enters free text (such as "I have often gotten lost in the past six months") into the electronic medical record interface, and the system calls the NLP model to extract the keyword "disorientation disorder".

[0066] Interference factors selection: If the structured checkbox contains options such as "arthritis / orthopedic disease" or "long-term use of benzodiazepines", selecting these options will trigger a data freeze rule.

[0067] Symptom Timeline: Integrates the Timeline JS 4.0 tool, allowing users to drag and drop to mark the start and end times of symptoms (e.g., "First time having difficulty naming in June 2023"), and automatically aligns them with the imaging examination time.

[0068] (2) Causal purification module, used to filter spurious correlations between multimodal data through clinical narrative anchoring and reverse intervention deduction, including:

[0069] Narrative anchoring unit: Based on a pre-defined causal conflict rule base, it identifies data conflicts through structured clinical narrative fields and triggers freeze or de-weighting operations;

[0070] Intervention simulation unit: Generates a reverse intervention validation package containing low-cost intervention measures and a review plan, and updates the causal rule base based on the review data;

[0071] The causal conflict rule base is constructed through a keyword mapping relationship table, where keywords include disease names, drug names, and life event keywords; when a keyword conflict is detected, the associated modal data is automatically frozen.

[0072] Specifically, the causal conflict rule base is constructed based on the MySQL table `causal_rules`, which defines keyword mapping relationships. For example, the keyword "benzodiazepines" is mapped to the genetic data APOE4, triggering a weight reduction operation (risk value × 0.5). The keyword "arthritis" is mapped to the behavioral data gait analysis, triggering a freeze operation (weight reset to zero). When clinical narrative information matches a keyword, the Drools rule engine automatically executes the operation and generates an audit log (JSON format recording patient ID, rule ID, and timestamp).

[0073] The reverse intervention verification package meets the following conditions:

[0074] The intervention does not conflict with the patient's current treatment plan;

[0075] The intervention cost is lower than the preset threshold;

[0076] The intervention effect is quantitatively evaluated using a formula:

[0077] Causal strength = (Δintervention group - Δcontrol group) / σbaseline.

[0078] Wherein, Δintervention group is the rate of change of the target modality index in the intervention group, Δcontrol group is the rate of change of the matched control group, and σbaseline is the baseline standard deviation; causal strength > 0.5 is marked as strong causality, 0.2-0.5 is marked as weak causality, and < 0.2 is marked as spurious correlation.

[0079] Intervention design includes:

[0080] Vitamin D supplementation: For patients with serum 25(OH)D <20 ng / mL, oral administration of 2000 IU / day is recommended (cost ≤50 yuan / month). The intervention effect is quantitatively verified by an improvement of ≥2 points in the MoCA score.

[0081] Sleep hygiene education: Distribute standardized manuals (including caffeine restriction recommendations), combine with weekly telephone follow-ups (labor cost ≤ 200 RMB / month), and evaluate the effectiveness through PSQI scoring.

[0082] Follow-up appointment plan generation: Based on historical data, the optimal follow-up appointment time is fitted and a standardized form is output (including a QR code link to remind the patient).

[0083] (3) Dynamic fusion module, used for time-aware alignment of time-series heterogeneous data, and modeling the interaction effect between pathological decline and behavioral compensation, including:

[0084] Time-varying coding unit: Captures the nonlinear relationship between brain atrophy rate and cognitive score in longitudinal multimodal data; the time-varying coding technique specifically involves inputting longitudinal MRI data (hippocampal volume sequence) and MoCA score, with a network structure of a 2-layer LSTM (128 hidden units, dropout=0.2), and outputting a prediction of the future atrophy rate (e.g., predicting that the hippocampal volume will decrease to 5.8 cm³ by 2025). The nonlinear relationship between brain atrophy rate and cognitive decline (RBF kernel function) is calculated, and a high-risk warning is triggered when the annual hippocampal atrophy rate >5% and the annual MoCA score decrease >3 points.

[0085] An explainable attention mechanism is employed: weighting the differentiation between pathological features and compensatory behaviors. Inputting a multimodal feature vector (images + genes + behavior), attention weights are used to distinguish between pathological features (e.g., Aβ deposition, weight > 0.7) and compensatory behaviors (e.g., electronic notepad usage, weight < 0.3). A heatmap is generated using the Grad-CAM algorithm, with red highlight areas showing the contribution of Aβ deposition (75%) and blue areas showing the contribution of compensatory behaviors (25%).

[0086] (4) Subtype classification module: a classification decision tree is constructed based on a biomarker framework, and a prototype network is used to handle the small sample learning problem of rare subtypes;

[0087] Specifically, the ATN framework classification decision tree is used.

[0088] The threshold is determined as follows:

[0089] A (Amyloid): Cerebrospinal fluid Aβ42 / Aβ40 < 0.08 or Aβ-PET SUVR > 1.2.

[0090] T(Tau): Cerebrospinal fluid p-Tau181 > 60 pg / mL or Tau-PET SUVR > 1.3.

[0091] N (Neurodegeneration): Hippocampal volume <6.5 cm³ (female) / <7.0 cm³ (male).

[0092] The classification logic is as follows:

[0093] A+T+N+ → Typical Alzheimer's disease (AD)

[0094] AT-N+ → Vascular cognitive impairment (VCI)

[0095] A+TN- → Primary progressive aphasia (PPA)

[0096] The prototype network learns from few samples, and the training process is as follows:

[0097] Support set: MRI features of 10 cases of posterior cortical atrophy AD (ResNet-50 extraction).

[0098] Query set: 5 cases to be classified, calculate the Euclidean distance to the prototype vector.

[0099] Classification results: Softmax output probability distribution, accuracy ≥ 85% (10-fold cross-validation).

[0100] (5) Clinical interaction module, outputting two-color label reports, causal quick lookup tools and standardized intervention verification suggestion forms.

[0101] The dual-color label report includes:

[0102] Red label: Mark frozen data (e.g., "Arthritis → Gait Analysis Frozen"). The HTML code is embedded in the electronic medical record check page. Clicking it will display the reason for freezing and processing suggestions.

[0103] Yellow label: Marks the association to be verified (e.g., "APOE4 positive but no family history"), AJAX dynamically loads evidence details (calls medication records and family history data in HIS).

[0104] Cause and effect quick reference tools include:

[0105] Tear-resistant posters: printed with high-frequency rules (such as "long-term depression → hippocampal volume diagnosis weight reduced by 50%), with a built-in NFC chip (reading distance ≤ 5cm), doctors can jump to the latest rule library when they bring their mobile phones close.

[0106] Standardized intervention card: includes checkbox options (such as vitamin D supplementation) and a follow-up appointment plan QR code. Patients scan the code and receive a notification on their mobile device (pushed 7 days, 3 days, or 1 day in advance).

[0107] An electronic medical record system integrates the aforementioned early screening and diagnosis system for cognitive impairment based on multimodal data, and interacts with the hospital's HIS system in the following ways:

[0108] Automatically extract medication records and comorbidity information from the HIS as clinical narrative fields;

[0109] Embed the two-color label report into the examination results page of the electronic medical record.

[0110] A computer-readable storage medium storing program instructions that, when executed by a processor, enable the function of the early screening and diagnosis system for cognitive impairment based on multimodal data.

[0111] A clinical toolkit for screening cognitive impairment includes:

[0112] A tear-resistant, cause-and-effect quick reference poster, printed with high-frequency conflict rules and with a reserved area for handwritten addition of new rules;

[0113] Red and yellow bicolor labels are used for manually marking conflicting data in inspection reports;

[0114] Standardized intervention verification cards, including check-in intervention options and QR codes linking to the follow-up plan.

[0115] A cognitive impairment screening method based on multimodal data includes the following steps:

[0116] (1) Collect multimodal data and clinical narrative information, and trigger data freezing through keyword mapping; for example, after selecting "history of knee replacement", the system freezes gait data and retains only hippocampal volume and gene data.

[0117] (2) Dynamically fuse the remaining data and model the temporal decay and compensation effects; for example, LSTM predicts that the annual decrease rate of hippocampal volume is >7%, and the attention mechanism shows that Aβ deposition contributes 80%.

[0118] (3) Output subtype classification results and two-color label report, and recommend intervention verification measures; if ATN is determined to be A+T+N+, output AD diagnosis report and vitamin D supplementation recommendations.

[0119] (4) Update the causal rule base and classification model based on the review data. For example, if the MoCA score improves by 3 points after the review, update "Vitamin D deficiency → cognitive improvement" to a strong causal relationship (causal strength = 1.2).

[0120] The following are case studies:

[0121] Patient: Ms. Zhang, 65 years old, chief complaint: "significant decline in memory in the past six months, often forgetting to turn off the gas", with a two-year history of taking lorazepam (benzodiazepine).

[0122] Step 1: Data Collection and Clinical Narrative Entry

[0123] Doctor: "Aunt Zhang, let's first record your symptoms and medical history. You mentioned that you've been forgetful lately, when exactly did this start?"

[0124] Ms. Zhang: "About half a year ago, at first I would occasionally forget where I put my keys. Last month, there were two near fires because I left the gas stove on."

[0125] Doctor's procedure: In the electronic medical record interface, check "long-term use of benzodiazepines" and enter the chief complaint text: "Memory decline in the past six months, and failure to turn off the gas twice."

[0126] The symptom timeline is marked as follows: "First memory loss in November 2023" → "Gas left unattended incident in April 2024".

[0127] The following tests were ordered: MRI hippocampal volume measurement, APOE gene testing, and Fitbit behavioral monitoring (wear for 7 days).

[0128] System response: Automatic rule base trigger: "Benzodiazepines" keyword detected, strong risk markers for the APOE4 gene frozen, log generated. A notification is pushed to the patient's mobile device: "Please go to the Radiology Department on the 3rd floor to complete the MRI examination."

[0129] Step 2: Data Integration and Causal Decomposition

[0130] The doctor reviewed the test results:

[0131] MRI report: hippocampal volume 6.1 cm³ (normal threshold for women >6.5 cm³), annual atrophy rate 6.2%.

[0132] Genetic report: APOE4 ε3 / ε4 positive, but marked with a yellow label due to drug conflict.

[0133] Behavioral data: Fitbit shows a gait symmetry index of 0.82 (normal >0.85), and an average of 3 awakenings per hour at night.

[0134] The system automatically processed the data: the dynamic fusion module predicted that, based on the current atrophy rate, the MoCA score might decrease from 24 to 20 after one year. Attention mechanism analysis showed Aβ-PET SUVR=1.1 (negative), but drug influence accounted for 60% of the weighting, while pathology contributed only 40%.

[0135] Step 3: Discussion of Diagnostic and Intervention Recommendations

[0136] Doctor: "Aunt Zhang, the system shows that your hippocampus is shrinking rapidly, but the genetic risk is affected by the medication and needs further verification. I suggest adjusting your sleeping pills first, and supplementing with vitamin D (detected value 18 ng / mL), and then having a follow-up examination after 3 months."

[0137] Ms. Zhang: "Will stopping the medication worsen my insomnia? How should I supplement with Vitamin D?"

[0138] Dr. Li's procedure: An intervention verification package was generated. The intervention involved gradually reducing the lorazepam dose (0.5 mg per week, under the supervision of a neurology consultation). Oral vitamin D3 2000 IU / day was prescribed. The follow-up plan included a MoCA score and serum vitamin D levels at month 3, and a hippocampal MRI at month 6. Standardized intervention cards were printed, and the patient scanned the code to sync the data to their mobile phone calendar.

[0139] System response: Insert a two-color label report into the electronic medical record:

[0140] Red label: "APOE4 risk weight freeze (lorazepam interference)".

[0141] Yellow label: "Hippocampal atrophy needs to rule out the influence of medication; follow-up examination is recommended."

[0142] Step 4: Patient Education and Tool Kit Use

[0143] Doctor: "This quick cause and effect check poster lists common confounding factors, such as long-term depression or arthritis, which can affect the diagnosis. If you have new symptoms, you can mark and report them with red and yellow stickers."

[0144] Ms. Zhang: "I understand. I'll put the vitamin D pillbox and the follow-up appointment reminder on the refrigerator when I get home."

[0145] Operation Demonstration: The doctor displays an anti-tear poster pointing to the rule "benzodiazepines → gene risk downweighting". The patient marks the "hippocampal atrophy" item in the MRI report with a yellow sticker and scans a QR code to view the follow-up guidelines.

[0146] Step 5: Review and rule base update

[0147] Follow-up examination 3 months later: MoCA score: increased from 24 to 26, sleep quality improved (PSQI decreased from 9 to 5). Serum vitamin D: increased to 32 ng / mL.

[0148] The system automatically analyzes:

[0149] Calculate the causal strength: Δintervention group (MoCA+2) - Δcontrol group (natural decline-1) / σbaseline (1.5) = (3) / 1.5 = 2.0 → strong causality

[0150] Update the rule base: Mark "Vitamin D supplementation → cognitive improvement" as a strong causal relationship (>0.5).

[0151] Doctor: "The results show that medication adjustments and vitamin D supplementation are effective, and the system has automatically optimized the diagnostic rules. We will continue to monitor and reassess the hippocampal volume in 6 months."

[0152] This case study demonstrates how the system addresses complex clinical issues such as drug interference and contradictory data through a process of structured data collection → dynamic causal reasoning → collaborative doctor-patient decision-making → closed-loop validation and updates. Doctors improve interpretation efficiency using tools such as two-color labels and intervention cards, while patients enhance compliance through visual reports and reminders, ultimately achieving personalized cognitive impairment management.

[0153] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A system for early screening and diagnosis of cognitive impairment based on multimodal data, characterized in that, Includes the following modules: (1) A multimodal data acquisition module, used to acquire image data, gene data, behavioral data and clinical narrative information; the clinical narrative information includes at least two of the following: patient complaint text, interfering factor checklist, and symptom timeline; (2) Causal purification module, used to filter spurious correlations between multimodal data through clinical narrative anchoring and reverse intervention deduction, including: Narrative anchoring unit: Based on a pre-defined causal conflict rule base, it identifies data conflicts through structured clinical narrative fields and triggers freeze or de-weighting operations; Intervention simulation unit: Generates a reverse intervention validation package containing low-cost intervention measures and a review plan, and updates the causal rule base based on the review data; (3) Dynamic fusion module, used for time-aware alignment of time-series heterogeneous data, and modeling the interaction effect between pathological decline and behavioral compensation, including: Time-varying coding unit: captures the nonlinear relationship between brain atrophy rate and cognitive score in longitudinal multimodal data; Explainable attention mechanisms: the weighting of pathological features and compensatory behaviors; (4) Subtype classification module: a classification decision tree is constructed based on a biomarker framework, and a prototype network is used to handle the small sample learning problem of rare subtypes; (5) Clinical interaction module, outputting two-color label reports, causal quick lookup tools and standardized intervention verification suggestion forms.

2. The system according to claim 1, characterized in that, The causal conflict rule base is constructed through a keyword mapping relationship table, where keywords include disease names, drug names, and life event keywords; when a keyword conflict is detected, the associated modal data is automatically frozen.

3. The system according to claim 1, characterized in that, The reverse intervention verification package meets the following conditions: The intervention does not conflict with the patient's current treatment plan; The intervention cost is lower than the preset threshold; The intervention effect is quantitatively evaluated using a formula: Causal strength = (Δintervention group - Δcontrol group) / σbaseline; Wherein, Δintervention group is the rate of change of the target modality index in the intervention group, Δcontrol group is the rate of change of the matched control group, and σbaseline is the baseline standard deviation; causal strength > 0.5 is marked as strong causation, causal strength between 0.2 and 0.5 is marked as weak causation, and causal strength < 0.2 is marked as spurious correlation.

4. The system according to claim 1, characterized in that, The two-color label report includes: Red label: Data frozen due to keyword conflicts; Yellow label: Marks associations that need to be verified.

5. The system according to claim 1, characterized in that, The cause-and-effect quick reference tool displays high-frequency conflict scenarios and handling rules in the form of interactive posters.

6. The system according to claim 1, characterized in that, The standardized intervention verification recommendation form includes checkbox options and a review plan template.

7. An electronic medical record system, characterized in that, Integrate the system according to any one of claims 1-6, and achieve the following interaction with the hospital HIS system: Automatically extract medication records and comorbidity information from the HIS as clinical narrative fields; Embed the two-color label report into the examination results page of the electronic medical record.

8. A computer-readable storage medium storing program instructions, characterized in that, When the program instructions are executed by the processor, they implement the functions of the system according to any one of claims 1-6.

9. A clinical toolkit for screening cognitive impairment, generated using a multimodal data-based early screening and diagnosis system for cognitive impairment as described in any one of claims 1-6, characterized in that, include: A tear-resistant, cause-and-effect quick reference poster, printed with high-frequency conflict rules and with a reserved area for handwritten addition of new rules; Red and yellow bicolor labels are used for manually marking conflicting data in inspection reports; Standardized intervention verification cards, including check-in intervention options and QR codes linking to the follow-up plan.

10. A cognitive impairment screening method based on multimodal data, applied to the early screening and diagnosis system for cognitive impairment based on multimodal data as described in any one of claims 1-6, characterized in that, Includes the following steps: (1) Collect multimodal data and clinical narrative information, and trigger data freezing through keyword mapping; (2) Dynamically fuse the remaining data and model the temporal decay and compensatory effects; (3) Output subtype classification results and two-color label report, and recommend intervention verification measures; (4) Update the causal rule base and classification model based on the review data.

Images