System and method for optimizing cognitive impairment intervention scheme of patient after ICU (intensive care unit)

Abstract

The invention discloses an optimization system and method for a cognitive disorder intervention scheme of a post-ICU patient, and belongs to the technical field of medical rehabilitation. The optimization system comprises a multi-dimensional evaluation module, a personalized intervention strategy generation module and a dynamic monitoring and feedback module; the multi-dimensional evaluation module comprises a cognitive function evaluation unit, a physiological parameter monitoring unit and a psychological state evaluation unit; the personalized intervention strategy generation module comprises a non-drug intervention library, a drug intervention decision tree and an environment optimization parameter library. Comprising the following steps: acquiring a dynamic data set through a multi-dimensional evaluation module; when the SpO2lt is detected; the MoCA directional force score is reduced by gt; when the concentration is 30%, generating a combined strategy of oxygen therapy and execution function training; and optimizing the intervention strategy weight based on the reinforcement learning model. The method has the advantages that a multi-dimensional dynamic evaluation system breaks through the limitation of traditional static evaluation; a closed-loop dynamic adjustment mechanism is adopted, so that accurate intervention is realized; a multi-dimensional collaborative intervention system breaks through a single treatment mode; and the decision-making accuracy is high.

Description

Technical Field

[0001] The present invention relates to the field of medical rehabilitation technology, and in particular to an optimization system and method for an intervention program for cognitive impairment in post-ICU patients. Background Art

[0002] Cognitive impairment in post-ICU patients, medically known as post-ICU syndrome (PICS-Cognitive), is a common complication in critically ill patients after discharge from the ICU and is one of the core manifestations of post-ICU syndrome (PICS). Domestic studies have shown that the incidence of cognitive impairment in post-ICU patients is as high as 44.7%, and some patients may experience symptoms for months or even years, significantly impacting their recovery and quality of life. Clinical manifestations of cognitive impairment in post-ICU patients include memory loss, difficulty concentrating, decreased executive function, impaired language skills, and visual-spatial impairments. Some patients may also experience delirium, anxiety, or depression.

[0003] Risk factors for cognitive impairment in post-ICU patients include:

[0004] 1. Disease-related factors:

[0005] Hypoxemia: Lack of oxygen in brain tissue leads to neuronal damage; electrolyte imbalance: such as hyponatremia and high blood sugar fluctuations; malnutrition: long-term bed rest leads to protein-energy depletion.

[0006] 2. Treatment-related factors:

[0007] Mechanical ventilation: endotracheal intubation causes anxiety, and the use of sedatives (such as propofol and midazolam) increases the risk of delirium; Antimicrobial drugs: Certain drugs may affect cognition through the blood-brain barrier.

[0008] 3. Environmental factors:

[0009] Closed management: The lack of an accompanying care system in the ICU exacerbates loneliness; continuous stimulation: noise, light, and frequent nursing operations interfere with sleep; sleep deprivation: nighttime treatment leads to biological clock disorders.

[0010] 4. Individual factors:

[0011] Age: Elderly patients have decreased brain reserve function; Education level: Those with low education have weaker cognitive recovery ability; Social support: Lack of family support has a poorer prognosis.

[0012] At present, intervention options for cognitive impairment in post-ICU patients include:

[0013] 1. Non-drug treatment:

[0014] 101. Cognitive rehabilitation training:

[0015] Memory training: using calendars and reminder notes to aid memory; attention training: number sorting, concentration games; language training: vocabulary practice, reading comprehension.

[0016] 102. Psychotherapy:

[0017] Supportive psychotherapy: Listening to the patient's feelings and providing emotional support; cognitive behavioral therapy: Correcting negative thought patterns, such as "I will never get better."

[0018] 103. Environmental intervention:

[0019] Sleep management: dim the lights at night and reduce noise; family participation: visit regularly every day to reduce loneliness.

[0020] 2. Drug treatment:

[0021] Cholinesterase inhibitors: such as donepezil, used for cognitive impairment related to Alzheimer's disease; drugs that improve cerebral circulation: such as nimodipine, used for vascular cognitive impairment; antipsychotic drugs: such as quetiapine, used with caution for severe hallucinations and delusions.

[0022] Post-ICU cognitive impairment is a complex problem involving multiple factors, requiring a comprehensive assessment of the patient's condition and the development of an individualized intervention plan. However, existing intervention plans have the following shortcomings:

[0023] Lack of individualization: Traditional methods rely on static assessments such as the MMSE / MoCA scale, which tend to overlook the correlation between various factors. For example, they cannot correlate multidimensional dynamic data such as hypoxemia (SpO2 <90%), environmental stress (nocturnal noise >55dB), and drug metabolism (propofol blood concentration). Manual assessments have a high rate of missed diagnoses, and elderly patients (>65 years old) have a higher rate of misdiagnosis.

[0024] Single function: The existing system only focuses on a single treatment method (such as drugs or cognitive training), does not integrate multi-dimensional measures such as environmental intervention and psychological support, lacks a coordination mechanism, and cannot achieve dynamic adjustment. Summary of the Invention

[0025] In order to solve the problems of the prior art, the present invention provides an optimization system and method for an intervention program for cognitive impairment in post-ICU patients.

[0026] In order to solve the above technical problems, the present invention is achieved through the following technical solutions: an optimization system and method for the intervention program of cognitive impairment in post-ICU patients.

[0027] An optimization system for intervention programs for cognitive impairment in post-ICU patients, including a multidimensional assessment module, a personalized intervention strategy generation module, and a dynamic monitoring and feedback module;

[0028] The multidimensional assessment module comprehensively quantifies the patient's cognitive, physiological, and psychological status through multimodal data collection and standardized assessment tools, providing an accurate baseline for intervention strategies. It includes a cognitive function assessment unit, a physiological parameter monitoring unit, and a psychological status assessment unit.

[0029] Based on the evaluation results, the personalized intervention strategy generation module constructs an intervention program library from three dimensions: non-drug, drug, and environment. It supports dynamic combination and priority adjustment, including a non-drug intervention library, a drug intervention decision tree, and an environmental optimization parameter library, among which:

[0030] The non-drug intervention library includes graded cognitive training games, VR directional training scenarios, and a personalized music therapy repertoire library;

[0031] The drug intervention decision tree uses patient risk factors as the root node, matches the medication regimen recommended by clinical guidelines, assesses the risk of drug interactions by calling the PharmGKB drug genome database, and outputs safety scores and dosage adjustment recommendations;

[0032] The environmental optimization parameter library stores environmental intervention standards based on evidence-based medicine, including ICU noise threshold, light intensity, and sleep cycle adjustment scheme;

[0033] The dynamic monitoring and feedback module generates risk scores by integrating multi-source data in real time, driving closed-loop adjustments to intervention strategies to ensure the timeliness and individualization of the plans.

[0034] Furthermore, the cognitive function assessment unit dynamically executes the MoCA / MMSE scale through voice interaction and touch screen, records the patient's operation response time and error type, the error type including memory error, calculation error, and logical error, and generates a cognitive function impairment profile including accuracy, reaction speed, and error pattern;

[0035] The physiological parameter monitoring unit collects heart rate variability (HRV), blood oxygen saturation (SpO2), and blood glucose fluctuation data through a wearable device, and synchronously connects to a bedside EEG monitoring device to collect the θ / α wave power ratio in the EEG monitoring device;

[0036] The psychological state assessment unit identifies anxiety / depression speech features based on the HADS scale and speech emotion analysis algorithm, and generates a psychological state heat map with time as the horizontal axis and emotion dimension scores as the vertical axis. The speech emotion analysis algorithm uses a convolutional neural network to extract speech rate, fundamental frequency, and loudness acoustic features.

[0037] Furthermore, the dynamic monitoring and feedback module includes:

[0038] Multi-source data fusion unit: The multi-source data fusion unit receives in real time the cognitive assessment error rate, the duration of SpO2 < 90%, and the frequency of nighttime noise > 55dB events;

[0039] Cognitive risk prediction model: Calculates a dynamic risk score based on the following formula:

[0040] Risk_score=0.3×MoCA_drop+0.2×HRV_std+0.25×(1-SpO2_avg)+0.15×

[0041] HADS_anxiety+0.1×noise_exceed

[0042] Among them, each parameter is standardized;

[0043] Closed-loop control unit: When Risk_score>0.7, VR relaxation training + light adjustment combined intervention is automatically triggered.

[0044] Furthermore, the EEG monitoring uses a lightweight 4-channel head-mounted device to detect delirium tendencies through the α / θ wave power ratio. When the ratio is <1.5 for 10 minutes, an early warning signal is generated.

[0045] Furthermore, the drug decision tree includes a drug interaction assessment submodule, which calls the PharmGKB database to verify the correlation between the CYP2B6 genotype and the propofol metabolism rate. The specific logic is: if the patient's CYP2B6 genotype is a slow metabolizer, the system automatically reduces the recommended propofol dose by 20% and prompts to monitor the risk of respiratory depression; if a drug metabolized by CYP3A4 is used in combination, the dose is adjusted and the dosing interval is extended by 2 hours.

[0046] Furthermore, the optimization system is provided with a family collaboration terminal, through which the patient's cognitive training progress report and environment optimization suggestions are pushed.

[0047] A method for optimizing an intervention program for cognitive impairment in post-ICU patients, comprising the following steps:

[0048] S1: Acquire a dynamic data set through the multi-dimensional evaluation module;

[0049] S2: When SpO2 < 90% and MoCA orientation score decreased by > 30%, a combined strategy of oxygen therapy + executive function training was generated;

[0050] S3: Optimize the intervention strategy weights based on the reinforcement learning model, and the reward function is:

[0051] Reward = ΔMoCA + 0.5 × sleep quality improvement - 0.3 × drug side effect index;

[0052] Among them, ΔMoCA is the change in MoCA score before and after intervention;

[0053] The improvement of sleep quality was assessed by the PSQI scale, where sleep quality improvement = PSQI before intervention - PSQI after intervention, ranging from 0 to 21;

[0054] The drug side effect index is graded according to the CTCAE classification: grade 1 = mild, index = 1; grade 2 = moderate, index = 2; and so on.

[0055] Furthermore, step S1 specifically includes: performing the MoCA / MMSE scale every morning to collect a cognitive baseline, collecting heart rate variability (HRV) and blood oxygen saturation (SpO2) through wearable devices every 2 hours, performing HADS interviews + voice emotion analysis every afternoon to collect psychological state data, and synchronously receiving α / θ wave power ratio data from EEG monitoring, ultimately forming a multi-source heterogeneous dataset with timestamps, in which structured data is stored in an SQL database and unstructured voice / image data is stored in an object storage system.

[0056] Furthermore, for step S2, when SpO2 < 90% and the MoCA orientation score decreased by > 30%, a combined oxygen therapy + executive function training strategy was generated, where the oxygen therapy plan was nasal cannula oxygen inhalation at 2 L / min, with a target SpO2 ≥ 95%; executive function training selected the intermediate Stroop task, twice a day, 10 minutes each time, for 3 days; at the same time, the environmental optimization parameter library was called to reduce the nighttime light intensity to 30 lux to improve sleep continuity.

[0057] Furthermore, the reinforcement learning model adopts a deep Q-network (DQN), the state space is defined as the current risk score, cognitive / physiological / psychological state indicators, and the action space is defined as optional intervention strategies. The strategy weights are optimized through simulated interaction with real clinical scenarios, and ultimately an individualized optimal intervention plan is output.

[0058] The advantages of the present invention compared with the prior art are:

[0059] 1. A multi-dimensional dynamic assessment system breaks through the limitations of traditional static assessments: By integrating multimodal data on cognitive function, physiological parameters, and psychological state, a dynamic assessment model is constructed. Compared to traditional static methods that rely on single scale assessments, this system can capture dynamic correlation data such as hypoxemia, environmental stress, and drug metabolism, significantly reducing the rate of missed diagnoses and the rate of misdiagnosis in elderly patients.

[0060] 2. Closed-loop dynamic adjustment mechanism for precise intervention: Based on a risk scoring model and reinforcement learning algorithm, the system can generate personalized intervention strategies in real time. For example, when SpO2 is <90% and the MoCA orientation score drops >30%, oxygen therapy (2L / min) and executive function training (intermediate Stroop task) are automatically triggered, and nighttime lighting is simultaneously adjusted to 30 lux to improve sleep continuity. Compared with traditional fixed plans, dynamic adjustment significantly improves cognitive recovery efficiency;

[0061] 3. A multi-dimensional collaborative intervention system breaks through the single treatment model: integrating non-drug, drug, and environmental interventions. For example, the drug decision tree verifies the association between CYP2B6 genotype and propofol metabolism through the PharmGKB database, automatically reducing the dose by 20% for patients with slow metabolization to reduce the risk of respiratory depression. The environmental optimization parameter library is linked to smart devices to reduce the peak nighttime noise level from 65dB to below 45dB, significantly improving sleep efficiency.

[0062] 4. Lightweight smart devices and efficient algorithm support: A 4-channel EEG headset is used for delirium early warning, which is 60% more portable than traditional 16-channel devices. The reinforcement learning model optimizes policy weights using the DQN algorithm. The state space covers 12-dimensional risk indicators, and the action space defines eight intervention combinations, achieving high decision-making accuracy.

[0063] 5. Deep family participation mechanism to build a full-cycle support network: The family collaboration terminal provides training progress reports and environment optimization suggestions. Compared with the traditional one-way education model, family participation is enhanced and patient rehabilitation compliance is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a hierarchical architecture diagram of an optimization system for a cognitive impairment intervention program for post-ICU patients according to the present invention.

[0065] Figure 2 This is a module internal architecture diagram of an optimization system for a cognitive impairment intervention program for post-ICU patients according to the present invention.

[0066] Figure 3 This is a full-process timing diagram of an optimization system and method for an intervention program for cognitive impairment in post-ICU patients according to the present invention.

[0067] Figure 4 This is a logic block diagram of error type identification of an optimization system and method for a cognitive impairment intervention program for post-ICU patients according to the present invention.

[0068] Figure 5 This is a multi-source data fusion timing diagram of an optimization system and method for an intervention program for cognitive impairment in post-ICU patients according to the present invention.

[0069] Figure 6This is a multimodal evaluation flow chart of an optimization system and method for an intervention program for cognitive impairment in post-ICU patients according to the present invention.

[0070] Figure 7 This is an acoustic feature extraction pipeline diagram of an optimization system and method for an intervention program for cognitive impairment in post-ICU patients according to the present invention.

[0071] Figure 8 The present invention provides a drug intervention decision tree implementation flow chart for an optimization system and method for a cognitive impairment intervention program for post-ICU patients.

[0072] Figure 9 The present invention provides a drug intervention decision tree decision logic diagram of an optimization system and method for a cognitive impairment intervention program for post-ICU patients. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0074] Example 1, combined with the attached Figure 1-9 :

[0075] An optimization system for intervention programs for cognitive impairment in post-ICU patients, including a multidimensional assessment module, a personalized intervention strategy generation module, and a dynamic monitoring and feedback module;

[0076] The multi-dimensional assessment module comprehensively quantifies patients' cognitive, physiological, and psychological states through multimodal data collection and standardized assessment tools, providing an accurate baseline for intervention strategies. It includes cognitive function assessment units, physiological parameter monitoring units, and psychological state assessment units.

[0077] The cognitive function assessment unit dynamically executes the MoCA / MMSE scale through voice interaction and a touch screen, recording the patient's operation response time and error types, including memory errors, calculation errors, and logical errors, and generates a cognitive impairment profile that includes accuracy, reaction speed, and error patterns.

[0078] The cognitive function assessment unit integrates a real-time clock (RTC) chip to achieve millisecond-level precision recording of operation response time; the voice interaction module uses a directional microphone array + environmental noise reduction chip, where the signal-to-noise ratio of the environmental noise reduction chip is >60dB; the touch screen module is equipped with a medical-grade anti-glare display with a pressure-sensitive layer, a sampling rate of ≥120Hz, and supports gesture operations such as sliding / clicking.

[0079] Implementation process:

[0080] 1. Voice interaction uses a text-to-speech (TTS) engine to output voice commands in the order of the questions, such as "Please click on the cube in the picture," and simultaneously triggers a timer. The touch response data is used to establish a mapping relationship between screen coordinates and question elements, such as: in the MoCA visual-spatial task, the cube vertices = coordinate areas P1-P8.

[0081] 2. Key parameter collection

[0082] Data Type Collection method Accuracy requirements Response time Time difference from the end of voice playback to the first touch ±10ms Operation Path Record the touch point coordinate sequence to form a trajectory map Resolution 0.1mm Error type tag Predefined error classification logic tree Classification accuracy > 95%

[0083] 3. Error type identification

[0084] Feature extraction methods

[0085] Memory errors: The number of times the wrong option was repeated in the “word recall” task was recorded, e.g., choosing “apple” instead of the target word “banana” three times;

[0086] Calculation error: Detect the deviation of the results in the "100 continuous subtraction 7" task, such as: 93→85→78→70 correct chain vs. 93→86→79 incorrect chain;

[0087] Logical error: Analyzing the pointer angle error in the "Clock Drawing" task, such as: the angle deviation between the hour hand and the minute hand is >10°.

[0088] 4. Generate a cognitive impairment profile

[0089] Multi-dimensional indicator fusion model:

[0090]

[0091] Accuracy: number of correct questions / total number of questions × 100%;

[0092] Reaction Speed: median response time for each question (excluding outliers);

[0093] Error Pattern Vector: [memory error frequency, computational error magnitude, logic error complexity].

[0094] 5. Generate radar chart and overlay time series for visual output:

[0095] Radar chart: displays the Z-score standardized scores of memory / calculation / logic dimensions;

[0096] Time series: A trend line showing response speed over a series of assessments, such as a response time curve for daily assessments.

[0097] The physiological parameter monitoring unit collects heart rate variability (HRV), blood oxygen saturation (SpO2), and blood glucose fluctuation data through wearable devices, and simultaneously connects to bedside EEG monitoring equipment to collect the θ / α wave power ratio in the EEG monitoring equipment;

[0098] Implementation process:

[0099] 1. Wearable devices

[0100] parameter Technical Solution Medical Certification HRV Medical-grade PPG sensor (wavelength 530nm / 940nm dual light source) ISO80601-2-61 <![CDATA[SpO2]]> Reflective blood oxygen module (sampling rate 100Hz) FDA510(k)ClassII Blood sugar fluctuations Continuous glucose monitor (indirect measurement of subcutaneous tissue fluid) CEMarkClassIIb

[0101] Synchronization mechanism: The clocks of three devices are synchronized through the Medical Internet of Things protocol (IEEE 11073PHD), with time drift ≤ 10ms.

[0102] 2.EEG monitoring equipment

[0103] Electrode layout: Fp1 / Fp2 (frontal lobe), O1 / O2 (occipital lobe), simplified version of the international 10-20 system.

[0104] Key parameter: Theta band (4-8Hz) / α band (8-13Hz) power ratio calculation formula:

[0105]

[0106] 3. Data Stream Collaborative Processing

[0107] Step 1: Real-time signal acquisition

[0108] parameter Acquisition frequency Transport Protocol <![CDATA[HRV / SpO2]]> 250ms / time BLE5.0 (AES encryption) blood sugar 5 minutes / time Wi-Fi Direct EEG raw data 200Hz Isolating the CAN bus

[0109] Step 2: Key feature extraction

[0110] HRV time domain analysis:

[0111] RMSSD (root mean square difference between adjacent RR intervals) calculation window: 5-minute sliding window

[0112] Example: RR sequence [800ms, 850ms, 820ms] → ΔRR = [50ms, -30ms] → RMSSD = √((50 2 +(-30)2) / 2)=41.2ms;

[0113] SpO2 abnormality detection:

[0114] Trigger condition: SpO2 < 90% for 3 consecutive sampling points and duration > 30 seconds;

[0115] Theta / alpha wave power ratio:

[0116] Sliding FFT was used with a 2-second window length, 50% overlap, and the ratio was updated every 10 seconds.

[0117] Step 3: Multi-source data fusion.

[0118] The Mental State Assessment Unit uses the HADS scale and a speech emotion analysis algorithm to identify anxious / depressed speech characteristics and generate a mental state heat map with time as the horizontal axis and emotion scores as the vertical axis. The speech emotion analysis algorithm uses a convolutional neural network to extract acoustic features of speech rate, fundamental frequency, and loudness. Anxiety is identified by a decrease in speech rate of >20% and a fluctuation in fundamental frequency of <50Hz, accompanied by an increase in sentence interruptions. Depression is identified by a uniform speech rate but a decrease in volume of >30%, with fewer interrogative sentences.

[0119] Implementation process:

[0120] 1. Digital implementation of the HADS scale

[0121] 101.Hardware Configuration

[0122] Medical noise-canceling microphone: frequency response 50Hz-16kHz (±3dB), equipped with a pop filter;

[0123] 7-inch touch screen: resolution 1920×1200, supports glove touch mode;

[0124] Privacy protection design: local storage encryption, real-time desensitization of audio data;

[0125] 102.Interaction Process

[0126] a. System startup: Automatically play guidance voice ("Now start the emotion assessment, please answer the following questions...")

[0127] b.Topic presentation:

[0128] - Text questions: The screen displays HADS questions (e.g. "I feel nervous or in pain")

[0129] - Voice questions: TTS announces the questions (for visually impaired patients)

[0130] c. Response collection:

[0131] - Touch selection: 4-level Likert scale button (0-3 points)

[0132] - Voice response: recordings of open-ended questions (e.g., "Describe your sleep last night")

[0133] d. Timer: Each question is limited to 30 seconds. If the time exceeds the limit, it will be automatically recorded as "unanswered".

[0134] 103.Data Standardization

[0135] The raw scores were converted into standard scores of 0-21;

[0136] Anxiety / depression subscales are calculated independently:

[0137] Anxiety subscale = Q1+Q3+Q5+Q7+Q9+Q11+Q13

[0138] Depression subscale = Q2+Q4+Q6+Q8+Q10+Q12+Q14

[0139] 2. Voice emotion realization

[0140] 201.Acoustic feature extraction pipeline is as follows Figure 7 As shown;

[0141] 202.Key characteristic parameters

[0142] feature Extraction method Analytical indicators speaking speed Voice Activity Detection (VAD) Segmentation Syllables / second, rate of change relative to baseline Fundamental frequency (F0) Autocorrelation method + parabolic interpolation Mean, standard deviation, and fluctuation range Loudness A-weighted sound pressure level calculation Energy envelope curve Statement Interruption Silence segment detection (>300ms is a valid interruption) Interruptions / minute Question sentence ratio Detection of rising fundamental frequency at the end of a sentence (ΔF0>35Hz) Number of interrogative sentences / total number of sentences

[0143] 203. Anxiety / Depression Identification Logic

[0144] Anxiety Trait Model:

[0145] Condition 1: Current speaking speed < baseline speaking speed × 0.8

[0146] Condition 2: Fundamental frequency standard deviation < 50Hz

[0147] Condition 3: Statement interruption frequency > baseline value + 2 times / minute

[0148] All conditions are met → Anxiety probability > 85%

[0149] Depression trait model:

[0150] Condition 1: Speech rate variation coefficient < 0.15 (uniform speech rate)

[0151] Condition 2: Average loudness < baseline value × 0.7

[0152] Condition 3: The proportion of interrogative sentences decreases by >40%

[0153] All conditions are met → Depression probability > 80%

[0154] 204.CNN Acoustic Analysis Architecture

[0155] Input layer: 40-dimensional Mel-spectrogram (time × frequency)

[0156] Convolutional layer: Conv1: 32 5×5 filters → ReLU activation

[0157] Conv2: 64 3×3 filters → ReLU activation

[0158] Pooling layer: Max pooling 2×2

[0159] Fully connected layer: 128 neurons → Dropout (0.5)

[0160] Output layer: Softmax classification (anxiety / depression / neutral)

[0161] 3. Generating a psychological state heat map

[0162] 301.Data fusion algorithm:

[0163] \text{Emotion Score}_t=w_1\times\text{HADS}_t+w_2\times\text{Voice}_t;

[0164] Where: w1 = 0.6 (scale weight)

[0165] w2 = 0.4 (speech analysis weight)

[0166] Time resolution: Update every 15 minutes.

[0167] 302. Heatmap Construction

[0168] Dimensions Implementation Visual Coding Timeline 24-hour segment (15-minute unit) Horizontal axis (0:00-23:45) Emotional intensity Normalized score (0-10 points) Color depth (blue→yellow→red) Anxiety / depression distinction Dual-channel overlay display Anxiety = red channel, depression = blue channel

[0169] 303. Output Example

[0170] Time:2025-06-13 14:30

[0171] Anxiety score: 6.2 / 10 (HADS: 5.8 + Voice: 2.4)

[0172] Depression score: 4.1 / 10 (HADS: 3.2 + Voice: 0.9)

[0173] Heat map unit: [x=58,y=6.2,color=#FFA500] (anxiety)

[0174] [x=58,y=4.1,color=#1E90FF](depression).

[0175] Based on the evaluation results, the personalized intervention strategy generation module constructs an intervention plan library from three dimensions: non-drug, drug, and environment. It supports dynamic combination and priority adjustment, including a non-drug intervention library, a drug intervention decision tree, and an environmental optimization parameter library.

[0176] The non-pharmacological intervention library includes graded cognitive training games (such as elementary memory matching [4×4 cards], intermediate Stroop tasks [color-word conflict], and advanced logical reasoning tasks), VR orientation training scenarios (virtual ward environments, patient-familiar scenes [such as home living rooms], and support for head tracking interaction), and a personalized music therapy repertoire library (selected based on the patient's previous music preferences, with a tempo of 60-80 BPM, moderate melodic complexity, and avoiding strong rhythmic stimulation).

[0177] The drug intervention decision tree uses patient risk factors as its root node and matches the medication regimen recommended by clinical guidelines, such as oxiracetam for cognitive improvement and dexmedetomidine for sedation. Patient risk factors include APACHE II score, number of comorbidities, and liver and kidney function indicators. The tree also assesses drug interaction risks, such as the risk of respiratory depression associated with the combination of propofol and remifentanil, by invoking genotype data for metabolic enzymes such as CYP2B6 and CYP3A4 from the PharmGKB drug genomic database. The tree then outputs a safety score and dosage adjustment recommendations, with the safety score ranging from 1 to 5.

[0178] Implementation process:

[0179] 1. Multi-source data collection

[0180] Risk factors Data Source Treatment APACHE II score Electronic Medical Records (EMR) Automatic extraction + manual verification Number of comorbidities ICD-10 Coding Statistics Natural Language Processing (NLP) recognition Liver and kidney function indicators Inspection System (LIS) Real-time API integration Genotype data PharmGKB database HL7FHIR standard interface call

[0181] 2. Risk Stratification Model

[0182] Risk level = 0.4 × APACHE_II + 0.3 × comorbidity index + 0.3 × liver and kidney function coefficient

[0183] Among them: -APACHE_II: normalized to the range of 0-1 (≥25 points = 1.0)

[0184] -Comorbidity index: 0.2 for each chronic disease (diabetes / heart failure / COPD, etc.)

[0185] -Liver and kidney function coefficient:

[0186] Liver function: Child-Pugh A = 0.2, B = 0.5, C = 1.0

[0187] Renal function: eGFR ≥ 90 = 0.1, 60-89 = 0.3, < 60 = 0.8

[0188] 3. Clinical rule base

[0189] Indications Drug of choice Second-choice drugs Contraindications Delirium sedation Dexmedetomidine Propofol Severe bradycardia Improved cognitive function Oxiracetam Donepezil History of epilepsy Acute pain Remifentanil Fentanyl intracranial hypertension

[0190] 4. Key metabolic enzyme processing logic

[0191]

[0192]

[0193] 5. Drug Interaction Matrix

[0194]

[0195] 6. Safety Score

[0196] 601. Scoring Algorithm:

[0197] Safety score = basic risk + genetic risk + physiological risk

[0198] Baseline risk: Evidence level (1-3 points)

[0199] Genetic risk: PM = 2 points, IM = 1 point, EM = 0 points

[0200] Physiological risk: Liver damage = 1.5 points, kidney damage = 1 point, hypoxia = 0.5 points per episode. Example: Propofol (basic risk 2.0) for a CYP2B6 PM (+2) with liver damage (+1.5) → total score 5.5602. Graded response mechanism

[0201] score Risk Level System response 1.0-2.0 Low risk Routine monitoring 2.1-3.5 Medium risk Yellow warning + dosage adjustment recommendation 3.6-5.0 High risk Red blocking + mandatory consultation request

[0202] 603. Dosage Adjustment Rules

[0203] IF safety score>3.0THEN

[0204] Adjustment range = (score - 3.0) / 2 * 100%

[0205] The maximum reduction is no more than 40%

[0206] Example: Rating 4.0 → Adjustment range = (4-3) / 2*100% = 50% → Actual reduction is 40%.

[0207] The environmental optimization parameter library stores environmental intervention standards based on evidence-based medicine, including ICU noise thresholds: <45 decibels during the day and <40 decibels at night, monitored in real time by a sound level meter; light intensity: 200-500 lux in the daytime activity area and <50 lux in the nighttime sleep area, using smart dimming LED lamps; and sleep cycle adjustment plans: such as a fixed bedtime [22:00-24:00], a nighttime wake-up interval ≥2 hours, and white noise assistance;

[0208] The dynamic monitoring and feedback module generates risk scores by integrating multi-source data in real time, driving closed-loop adjustments to intervention strategies to ensure the timeliness and individualization of the plans.

[0209] The dynamic monitoring and feedback module includes:

[0210] Multi-source data fusion unit: The multi-source data fusion unit receives in real time the cognitive assessment error rate, SpO2<90% duration, and frequency of nighttime noise>55dB events;

[0211] Cognitive risk prediction model: Calculates a dynamic risk score based on the following formula:

[0212] Risk_score=0.3×MoCA_drop+0.2×HRV_std+0.25×(1-SpO2_avg)+0.15×

[0213] HADS_anxiety+0.1×noise_exceed

[0214] Among them, each parameter is standardized;

[0215] Closed-loop control unit: When Risk_score>0.7, VR relaxation training + light adjustment combined intervention is automatically triggered.

[0216] EEG monitoring uses a lightweight 4-channel headset to detect delirium tendencies through the α / θ wave power ratio. When the ratio is <1.5 for 10 minutes, an early warning signal is generated.

[0217] The drug decision tree includes a drug interaction assessment submodule, which calls the PharmGKB database to verify the correlation between the CYP2B6 genotype and the propofol metabolism rate. The specific logic is: if the patient's CYP2B6 genotype is a slow metabolizer, the system automatically reduces the recommended propofol dose by 20% and prompts monitoring for respiratory depression risk; if drugs metabolized by CYP3A4 are used in combination, the dose will be adjusted and the dosing interval will be extended by 2 hours.

[0218] The optimization system is equipped with a family collaboration terminal, through which patient cognitive training progress reports and environmental optimization suggestions are pushed.

[0219] A method for optimizing an intervention program for cognitive impairment in post-ICU patients, comprising the following steps:

[0220] S1: Obtain dynamic datasets through a multi-dimensional evaluation module;

[0221] S2: When SpO2 < 90% and MoCA orientation score decreased by > 30%, a combined strategy of oxygen therapy + executive function training was generated;

[0222] S3: Optimize the intervention strategy weights based on the reinforcement learning model, and the reward function is:

[0223] Reward = ΔMoCA + 0.5 × sleep quality improvement - 0.3 × drug side effect index;

[0224] Wherein, ΔMoCA is the change in MoCA score before and after the intervention, for example, if the score is +3 after the intervention, then ΔMoCA = 3;

[0225] The improvement of sleep quality was assessed by the PSQI (Pittsburgh Sleep Quality Index), where sleep quality improvement = PSQI before intervention - PSQI after intervention, ranging from 0 to 21.

[0226] The drug side effect index is graded according to the CTCAE classification: grade 1 = mild, index = 1; grade 2 = moderate, index = 2; and so on.

[0227] Step S1 specifically includes: performing the MoCA / MMSE scale every morning to collect a cognitive baseline, collecting heart rate variability (HRV) and blood oxygen saturation (SpO2) through wearable devices every 2 hours, performing the HADS interview and voice emotion analysis every afternoon to collect psychological state data, and synchronously receiving α / θ wave power ratio data from EEG monitoring. Ultimately, a multi-source heterogeneous dataset with timestamps is formed, where structured data is stored in an SQL database and unstructured voice / image data is stored in an object storage system.

[0228] In step S2, when SpO2 < 90% and the MoCA orientation score decreased by > 30%, a combined oxygen therapy + executive function training strategy was generated, where the oxygen therapy plan was nasal cannula oxygen inhalation at 2 L / min, with a target SpO2 ≥ 95%. Executive function training selected the intermediate Stroop task, twice daily for 10 minutes each time, for 3 days. At the same time, the environmental optimization parameter library was called to reduce the nighttime light intensity to 30 lux to improve sleep continuity.

[0229] The reinforcement learning model uses a deep Q-network (DQN). The state space is defined as the current risk score and cognitive / physiological / psychological state indicators, and the action space is defined as optional intervention strategies, such as "VR training," "drug adjustment," and "environmental adjustment." The strategy weights are optimized through simulated interaction with real clinical scenarios, and the optimal individualized intervention plan is ultimately output. The simulated interaction with real clinical scenarios is achieved through historical data playback and partial real data online learning.

[0230] The present invention and its implementation methods are described above. This description is not restrictive. If ordinary technicians in this field are inspired by it and design embodiments similar to the technical solution without creatively designing them without departing from the purpose of the invention, they should all fall within the scope of protection of the present invention.

Claims

1. A system for optimizing intervention programs for cognitive impairment in post-ICU patients, characterized by: It includes a multi-dimensional assessment module, a personalized intervention strategy generation module, and a dynamic monitoring and feedback module; The multidimensional assessment module comprehensively quantifies the patient's cognitive, physiological, and psychological status through multimodal data collection and standardized assessment tools, providing an accurate baseline for intervention strategies. It includes a cognitive function assessment unit, a physiological parameter monitoring unit, and a psychological status assessment unit. Based on the evaluation results, the personalized intervention strategy generation module constructs an intervention program library from three dimensions: non-drug, drug, and environment. It supports dynamic combination and priority adjustment, including a non-drug intervention library, a drug intervention decision tree, and an environmental optimization parameter library, among which: The non-drug intervention library includes graded cognitive training games, VR directional training scenarios, and a personalized music therapy repertoire library; The drug intervention decision tree uses patient risk factors as the root node, matches the medication regimen recommended by clinical guidelines, assesses the risk of drug interactions by calling the PharmGKB drug genome database, and outputs safety scores and dosage adjustment recommendations; The environmental optimization parameter library stores environmental intervention standards based on evidence-based medicine, including ICU noise threshold, light intensity, and sleep cycle adjustment scheme; The dynamic monitoring and feedback module generates risk scores by integrating multi-source data in real time, driving closed-loop adjustments to intervention strategies to ensure the timeliness and individualization of the plans.

2. The optimization system for the intervention program for cognitive impairment in post-ICU patients according to claim 1, characterized in that: The cognitive function assessment unit dynamically executes the MoCA / MMSE scale through voice interaction and a touch screen, records the patient's operation response time and error types, including memory errors, calculation errors, and logical errors, and generates a cognitive function impairment profile that includes accuracy, reaction speed, and error patterns; The physiological parameter monitoring unit collects heart rate variability (HRV), blood oxygen saturation (SpO2), and blood glucose fluctuation data through a wearable device, and synchronously connects to a bedside EEG monitoring device to collect the θ / α wave power ratio in the EEG monitoring device; The psychological state assessment unit identifies anxiety / depression speech features based on the HADS scale and speech emotion analysis algorithm, and generates a psychological state heat map with time as the horizontal axis and emotion dimension scores as the vertical axis. The speech emotion analysis algorithm uses a convolutional neural network to extract speech rate, fundamental frequency, and loudness acoustic features.

3. The optimization system for the intervention program for cognitive impairment in post-ICU patients according to claim 1, characterized in that: The dynamic monitoring and feedback module includes: Multi-source data fusion unit: The multi-source data fusion unit receives in real time the cognitive assessment error rate, the duration of SpO2 < 90%, and the frequency of nighttime noise > 55dB events; Cognitive risk prediction model: Calculates a dynamic risk score based on the following formula: Risk_score=0.3×MoCA_drop+0.2×HRV_std+0.25×(1-SpO2_avg)+0.15× HADS_anxiety+0.1×noise_exceed Among them, each parameter is standardized; Closed-loop control unit: When Risk_score>0.7, VR relaxation training + light adjustment combined intervention is automatically triggered.

4. The optimization system for the intervention program for cognitive impairment in post-ICU patients according to claim 1, characterized in that: The EEG monitoring uses a lightweight 4-channel head-mounted device to detect delirium tendencies through the α / θ wave power ratio. When the ratio is <1.5 for 10 minutes, an early warning signal is generated.

5. The optimization system for intervention programs for cognitive impairment in post-ICU patients according to claim 1, characterized in that: The drug decision tree includes a drug interaction assessment submodule, which calls the PharmGKB database to verify the association between the CYP2B6 genotype and the propofol metabolism rate. The specific logic is: if the patient's CYP2B6 genotype is a slow metabolizer, the system automatically reduces the recommended propofol dose by 20% and prompts monitoring for respiratory depression risk; if a drug metabolized by CYP3A4 is used in combination, the dose is adjusted and the dosing interval is extended by 2 hours.

6. The optimization system for intervention programs for cognitive impairment in post-ICU patients according to claim 1, characterized in that: The optimization system is provided with a family collaboration terminal, through which the patient's cognitive training progress report and environment optimization suggestions are pushed.

7. A method for optimizing an intervention program for cognitive impairment in post-ICU patients, the method being based on the optimization system according to any one of claims 1 to 6, characterized in that: The following steps are involved: S1: Acquire a dynamic data set through the multi-dimensional evaluation module; S2: When SpO2 < 90% and MoCA orientation score decreased by > 30%, a combined strategy of oxygen therapy + executive function training was generated; S3: Optimize the intervention strategy weights based on the reinforcement learning model, and the reward function is: Reward = ΔMoCA + 0.5 × sleep quality improvement - 0.3 × drug side effect index; Among them, ΔMoCA is the change in MoCA score before and after intervention; The improvement of sleep quality was assessed by the PSQI scale, where sleep quality improvement = PSQI before intervention - PSQI after intervention, ranging from 0 to 21; The drug side effect index is graded according to the CTCAE classification: grade 1 = mild, index = 1; grade 2 = moderate, index = 2; and so on.

8. The method for optimizing an intervention program for cognitive impairment in post-ICU patients according to claim 7, characterized in that: Step S1 specifically includes: performing the MoCA / MMSE scale every morning to collect a cognitive baseline, collecting heart rate variability (HRV) and blood oxygen saturation (SpO2) through wearable devices every 2 hours, performing the HADS interview and voice emotion analysis every afternoon to collect psychological state data, and synchronously receiving α / θ wave power ratio data from EEG monitoring. Ultimately, a multi-source heterogeneous dataset with timestamps is formed, where structured data is stored in an SQL database and unstructured voice / image data is stored in an object storage system.

9. The method for optimizing an intervention program for cognitive impairment in post-ICU patients according to claim 7, characterized in that: In step S2, when SpO2 < 90% and the MoCA orientation score decreased by > 30%, a combined oxygen therapy + executive function training strategy was generated, where the oxygen therapy plan was nasal cannula oxygen inhalation at 2 L / min, with a target SpO2 ≥ 95%. Executive function training selected the intermediate Stroop task, twice daily for 10 minutes each time, for 3 days. At the same time, the environmental optimization parameter library was called to reduce the nighttime light intensity to 30 lux to improve sleep continuity.

10. The method for optimizing an intervention program for cognitive impairment in post-ICU patients according to claim 7, characterized in that: The reinforcement learning model uses a deep Q-network (DQN). The state space is defined as the current risk score and cognitive / physiological / psychological state indicators, and the action space is defined as the optional intervention strategies. The strategy weights are optimized through simulated interaction with real clinical scenarios, and the individualized optimal intervention plan is finally output.

Images