An ICU monitoring self-defined voice hierarchical alarm system and method based on personalized voiceprint

Abstract

This invention discloses a customized voice-based graded alarm system and method for ICU monitoring based on personalized voiceprints. It belongs to the field of medical monitoring device technology. The system includes: a data acquisition module for acquiring patients' vital sign data from a central monitoring system in real time; a risk grading module for grading the current risk according to predefined rules and / or machine learning models, outputting the risk level and corresponding warning information; a personalized voice synthesis module for generating broadcast text based on the risk level and warning information, and synthesizing the text into a speech signal using a preset specific human voiceprint model; and a broadcast control module for receiving the speech signal and controlling the playback device to broadcast it. This invention improves the recognition and urgency of alarms through dynamic risk assessment and personalized voice synthesis, thereby reducing alarm fatigue and improving the efficiency of medical staff response.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring device technology, and more specifically to an ICU monitoring customized voice hierarchical alarm system and method based on personalized voiceprints. Background Technology

[0002] The Intensive Care Unit (ICU) is the core location in a hospital for the centralized treatment of critically ill patients. To ensure patient safety, ICUs are typically equipped with a central monitoring system that continuously monitors multiple vital signs, including but not limited to heart rate, blood pressure, blood oxygen saturation, respiratory rate, and body temperature. These monitoring systems generally employ a tiered alarm mechanism: when one or more vital signs exceed a preset safety threshold, the system triggers different levels of audible and visual alarms based on the degree of deviation. For example, low-risk abnormalities trigger flashing yellow or blue lights and a soft beep, while life-threatening severe abnormalities trigger flashing red lights and a sharp, urgent alarm.

[0003] However, in clinical practice, this traditional tiered audible and visual alarm system is facing serious challenges. Due to the large number of monitoring devices in the ICU and the complex and ever-changing conditions of patients, alarm events occur extremely frequently. Statistics show that a single ICU bed can experience hundreds of alarms per day, many of which are non-emergency "technical alarms" (such as electrode detachment or loose blood pressure cuffs) or physiological abnormalities of little clinical significance. Prolonged exposure to a frequent and homogenized alarm environment gradually leads to "alarm fatigue" among medical staff, manifested as decreased sensitivity to alarm signals, delayed response, or even ignoring them. This problem has been listed as the number one medical technology hazard by the National Emergency Medical Research Institute (ECRI) for many consecutive years. Alarm fatigue directly threatens patient safety, potentially leading to delayed treatment of critical situations and serious consequences.

[0004] To alleviate alarm fatigue, existing technologies have attempted some improvements. For example, some monitoring systems allow adjusting alarm thresholds or setting delays to reduce false alarms; some studies have also proposed using machine learning algorithms to intelligently filter alarm signals and reduce false alarm rates. However, these solutions mainly focus on optimizing the "accuracy" of alarms, without fundamentally improving the "communication method" of alarm information. Healthcare workers still receive monotonous beeping sounds or flashing lights, lacking intuitive indications of the patient's specific condition and urgency.

[0005] In recent years, voice alarm technology has begun to attract attention. Some studies have attempted to introduce text-to-speech (TTS) technology into monitoring systems to directly announce brief information such as "patient's blood pressure has dropped." However, existing voice alarm solutions have the following significant shortcomings: Voice homogenization: The speech is mostly standard synthesized sound, lacking individual characteristics, and not much different from background noise, so it is still difficult to attract attention quickly.

[0006] Low risk differentiation: Alarms of different urgency levels often use the same voice style (such as the same speaking speed and tone), which cannot intuitively convey the sense of urgency through hearing.

[0007] The broadcast content is fixed: it consists mostly of preset fixed phrases and cannot dynamically generate personalized broadcasts that include specific patient information (such as bed number, name, and precise values) based on real-time data, resulting in limited information content.

[0008] Risk assessment relies on static thresholds: the fixed threshold rules of traditional monitoring systems are still used, which makes it difficult to capture trend changes in vital signs and results in early warning delays.

[0009] In summary, existing ICU monitoring alarm systems have significant shortcomings in their alarm transmission methods. There is an urgent need for an intelligent voice alarm solution that can effectively penetrate alarm fatigue, provide intuitive, personalized, and information-rich responses that can differentiate risk levels, thereby improving the response efficiency of medical staff and ensuring patient safety. Summary of the Invention

[0010] In view of this, the present invention provides a customized voice-based hierarchical alarm system and method for ICU monitoring based on personalized voiceprints. Through dynamic risk assessment and personalized voice synthesis, the system improves the recognition and urgency of alarms, thereby reducing alarm fatigue and improving the response efficiency of medical staff.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: A personalized voiceprint-based customized voice-based hierarchical alarm system for ICU monitoring includes: The data acquisition module is used to acquire patients' vital signs data from the central monitoring system in real time; The risk classification module, connected to the data acquisition module, is used to classify the current risk according to predefined rules and / or machine learning models, and output the risk level and corresponding early warning information. The personalized speech synthesis module, connected to the risk classification module, is used to generate broadcast text based on the risk level and warning information, and synthesize the text into a speech signal using a preset specific human voiceprint model. Different risk levels correspond to different speech synthesis parameters. The broadcast control module, connected to the personalized speech synthesis module, is used to receive voice signals and control the playback device to broadcast.

[0012] Furthermore, the risk classification module includes: The rules engine classifies real-time vital sign data based on preset thresholds; The neural network prediction unit uses a temporal neural network to perform dynamic trend analysis on multi-parameter vital sign data, outputs the predicted risk probability, and merges it with the results of the rule engine to obtain the final risk level.

[0013] Furthermore, the temporal neural network is a Long Short-Term Memory (LSTM) network, and its update formula is:

[0014] in, Let be the vital signs vector at time t. In hidden state, In cellular state, , , These are the forget gate, input gate, and output gate, respectively. , , , This is the weight matrix. , , , For bias terms, is the sigmoid function, and ⊙ represents element-wise multiplication; the hidden state at the last time step is input into the fully connected layer and then passed through softmax to obtain the probability of each risk level.

[0015] Furthermore, the personalized speech synthesis module includes: The text generation submodule generates broadcast text from a preset template library based on the risk level and warning information; The voiceprint feature extraction submodule stores or extracts the voiceprint feature vector of the target speaker; The speech synthesis submodule uses a neural network-based text-to-speech (TTS) model, which combines voiceprint feature vectors to synthesize the broadcast text into speech, and adjusts the speech rate and fundamental frequency parameters according to the risk level.

[0016] Furthermore, the speech synthesis submodule adopts the Tacotron 2 architecture, the vocoder uses WaveGlow, and speaker embedding vectors are introduced in the encoder-decoder.

[0017] Furthermore, the adjustment of speech rate and fundamental frequency parameters according to risk level includes:

[0018] in, For speaking speed, Based on basic speaking speed, For the fundamental frequency, Based on the fundamental frequency, , For adjustment coefficients, This is a quantified value for the risk level.

[0019] Furthermore, the vital signs data include at least one or more of blood oxygen saturation, heart rate, and blood pressure; the warning information includes bed number, patient identification, and abnormal vital signs values.

[0020] A method for customized voice-based graded alarms in ICU monitoring based on personalized voiceprints includes the following steps: S1: Real-time acquisition of patient vital signs data; S2: Classify the current risk according to predefined rules and / or machine learning models, and output the risk level and corresponding early warning information; S3: Generate broadcast text based on risk level and warning information, and synthesize the text into speech signal using a preset specific human voiceprint model, where different risk levels correspond to different speech synthesis parameters; S4: Broadcast the aforementioned voice signal.

[0021] As can be seen from the above technical solution, compared with the prior art, the present invention provides an ICU monitoring customized voice hierarchical alarm system and method based on personalized voiceprint. Through dynamic risk assessment, combined with rules and neural networks, it improves the accuracy and foresight of risk prediction and reduces missed and false alarms. Personalized voice synthesis uses voices familiar to medical staff, enhancing the affinity and authority of the alarm, while distinguishing risk levels through tone changes, effectively alleviating alarm fatigue. The system has a modular design, is easy to integrate with existing monitoring systems, and supports customized broadcast content and voice to adapt to the needs of different departments. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present 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 only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the system structure provided by the present invention; Figure 2 This is a schematic diagram of the method flow provided by the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1: See Figure 1 Embodiment 1 of the present invention discloses a customized voice-based hierarchical alarm system for ICU monitoring based on personalized voiceprints, comprising: The data acquisition module is used to acquire patients' vital signs data from the central monitoring system in real time; The risk classification module, connected to the data acquisition module, is used to classify the current risk according to predefined rules and / or machine learning models, and output the risk level and corresponding early warning information. The personalized speech synthesis module, connected to the risk classification module, is used to generate broadcast text based on the risk level and warning information, and synthesize the text into a speech signal using a preset specific human voiceprint model. Different risk levels correspond to different speech synthesis parameters. The broadcast control module, connected to the personalized speech synthesis module, is used to receive voice signals and control the playback device to broadcast.

[0026] In one specific embodiment, the risk classification module includes: The rules engine classifies real-time vital sign data based on preset thresholds; The neural network prediction unit uses a temporal neural network to perform dynamic trend analysis on multi-parameter vital sign data, outputs the predicted risk probability, and merges it with the results of the rule engine to obtain the final risk level.

[0027] In one specific embodiment, the temporal neural network is a Long Short-Term Memory (LSTM) network, and its update formula is:

[0028] in, Let be the vital signs vector at time t. In hidden state, In cellular state, , , These are the forget gate, input gate, and output gate, respectively. , , , This is the weight matrix. , , , For bias terms, is the sigmoid function, and ⊙ represents element-wise multiplication; the hidden state at the last time step is input into the fully connected layer and then passed through softmax to obtain the probability of each risk level.

[0029] In one specific embodiment, the personalized speech synthesis module includes: The text generation submodule generates broadcast text from a preset template library based on the risk level and warning information; The voiceprint feature extraction submodule stores or extracts the voiceprint feature vector of the target speaker; The speech synthesis submodule uses a neural network-based text-to-speech (TTS) model, which combines voiceprint feature vectors to synthesize the broadcast text into speech, and adjusts the speech rate and fundamental frequency parameters according to the risk level.

[0030] In one specific embodiment, the speech synthesis submodule adopts the Tacotron 2 architecture, the vocoder uses WaveGlow, and speaker embedding vectors are introduced in the encoder-decoder.

[0031] In one specific embodiment, adjusting the speech rate and fundamental frequency parameters according to the risk level includes:

[0032] in, For speaking speed, Based on basic speaking speed, For the fundamental frequency, Based on the fundamental frequency, , For adjustment coefficients, This is a quantified value for the risk level.

[0033] In one specific embodiment, the vital signs data include at least one or more of blood oxygen saturation, heart rate, and blood pressure; the warning information includes bed number, patient identifier, and abnormal vital signs values.

[0034] See Figure 2 Embodiment 1 of the present invention also discloses a method for customized voice-based graded alarms in ICU monitoring based on personalized voiceprints, comprising the following steps: S1: Real-time acquisition of patient vital signs data; S2: Classify the current risk according to predefined rules and / or machine learning models, and output the risk level and corresponding early warning information; S3: Generate broadcast text based on risk level and warning information, and synthesize the text into speech signal using a preset specific human voiceprint model, where different risk levels correspond to different speech synthesis parameters; S4: Broadcast the aforementioned voice signal.

[0035] Example 2: Embodiment 2 of the present invention discloses a customized voice-based hierarchical alarm system for ICU monitoring based on personalized voiceprints, comprising: Data acquisition module: used to acquire patients' vital signs data from the central monitoring system in real time, including one or more of the following: blood oxygen saturation, heart rate, and blood pressure.

[0036] Risk classification module: Connected to the data acquisition module, it is used to classify the current risk according to predefined rules and / or machine learning models, and output the risk level and corresponding early warning information. Specifically, the risk levels include low risk, medium risk, and high risk; The warning information includes: bed number, patient identification, and abnormal vital sign values.

[0037] Personalized speech synthesis module: Connected to the risk grading module, it generates broadcast text based on risk level and warning information, and synthesizes the text into speech signal using a preset specific human voiceprint model. Different risk levels correspond to different speech rate, pitch and intonation parameters.

[0038] Broadcast control module: Connected to the personalized speech synthesis module, it is used to receive voice signals and control the playback device to broadcast.

[0039] The risk classification module includes: Rule engine: Classifies real-time vital sign data based on preset thresholds; Neural Network Prediction Unit: Employs a temporal neural network to perform dynamic trend analysis on multi-parameter vital sign data, outputs the predicted risk probability, and integrates it with the results of the rule engine to obtain the final risk level.

[0040] The neural network prediction unit employs a Long Short-Term Memory (LSTM) network, and its calculation process is as follows: Let the input sequence be Where T is the length of the input sequence. LSTM processes the sequence sequentially at time steps t=1,2,...,T. Let be the n-dimensional vital sign vector at time t, which in this embodiment represents heart rate, blood oxygen saturation, and blood pressure. The update formula for the LSTM unit at time t is:

[0041] in, Let be the vital signs vector at time t. In hidden state, In cellular state, , , These are the forget gate, input gate, and output gate, respectively. , , , This is the weight matrix. , , , For bias terms, is the sigmoid function, and ⊙ represents element-wise multiplication; the hidden state at the last time step is input into the fully connected layer and then passed through softmax to obtain the probability of each risk level.

[0042] By introducing LSTM and clarifying its calculation process, the problems of alarm delay and low sensitivity in existing technologies are solved.

[0043] The personalized speech synthesis module includes: Text generation submodule: Based on the risk level and warning information, select the corresponding template from the preset template library, fill in the specific data, and generate the broadcast text; Voiceprint feature extraction submodule: Stores or extracts the voiceprint feature vector of the target speaker; The speech synthesis submodule uses a neural network-based text-to-speech (TTS) model to synthesize the broadcast text into speech. During the synthesis process, parameters such as duration and fundamental frequency are adjusted to achieve speech output with different levels of urgency.

[0044] The TTS model employs the Tacotron 2 architecture, with WaveGlow used for its vocoder. To achieve personalized voiceprints, a speaker embedding vector is introduced into the Tacotron 2 encoder-decoder structure. This vector is extracted from a small number of speech samples from the target speaker through a pre-trained speaker verification network (x-vector) and fine-tuned during training. The speech rate and fundamental frequency of the synthesized speech are dynamically adjusted according to the risk level.

[0045] in, For speaking speed, Based on basic speaking speed, For the fundamental frequency, Based on the fundamental frequency, , For adjustment coefficients, In this embodiment, the risk level is quantified as low risk = 0, medium risk = 1, and high risk = 2.

[0046] Embodiment 2 of the present invention also discloses a method for customized voice-based graded alarms in ICU monitoring based on personalized voiceprints, comprising the following steps: S1: Real-time acquisition of patient vital signs data; S2: Classify the current risk according to predefined rules and / or machine learning models, and output the risk level and corresponding early warning information; S3: Generate broadcast text based on risk level and warning information, and synthesize the text into speech signal using a preset specific human voiceprint model, where different risk levels correspond to different speech synthesis parameters; S4: Broadcast the aforementioned voice signal.

[0047] Specifically, most existing monitoring systems rely on fixed thresholds for alarms. These static rules struggle to capture the slow, gradual changes in vital signs, easily leading to missed or delayed alarms. This application introduces LSTM into ICU monitoring, where the LSTM input... It can simultaneously include vital sign data from multiple dimensions such as heart rate, blood oxygen, and blood pressure, through hidden states. By integrating the interactions between different parameters, complex physiological changes can be captured. This multi-parameter fusion capability far surpasses traditional single-parameter threshold alarms, enabling a more comprehensive assessment of the patient's condition and reducing false alarms.

[0048] Specifically, this application also integrates the prediction results of LSTM with those of the rule engine to obtain the final risk level. This dual-judgment mechanism leverages both the reliability of clinical experience and the sensitivity of LSTM machine learning, improving the accuracy of early warnings while ensuring safety.

[0049] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0050] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A customized voice-based graded alarm system for ICU monitoring based on personalized voiceprints, characterized in that, include: The data acquisition module is used to acquire patients' vital signs data from the central monitoring system in real time; The risk classification module, connected to the data acquisition module, is used to classify the current risk according to predefined rules and / or machine learning models, and output the risk level and corresponding early warning information. The personalized speech synthesis module, connected to the risk classification module, is used to generate broadcast text based on the risk level and warning information, and synthesize the text into a speech signal using a preset specific human voiceprint model. Different risk levels correspond to different speech synthesis parameters. The broadcast control module, connected to the personalized speech synthesis module, is used to receive voice signals and control the playback device to broadcast.

2. The ICU monitoring customized voice hierarchical alarm system based on personalized voiceprint as described in claim 1, characterized in that, The risk classification module includes: The rules engine classifies real-time vital sign data based on preset thresholds; The neural network prediction unit uses a temporal neural network to perform dynamic trend analysis on multi-parameter vital sign data, outputs the predicted risk probability, and merges it with the results of the rule engine to obtain the final risk level.

3. The ICU monitoring customized voice hierarchical alarm system based on personalized voiceprint as described in claim 2, characterized in that, The temporal neural network is a Long Short-Term Memory (LSTM) network, and its update formula is: in, Let be the vital signs vector at time t. In hidden state, In cellular state, , , These are the forget gate, input gate, and output gate, respectively. , , , This is the weight matrix. , , , For bias terms, is the sigmoid function, and ⊙ represents element-wise multiplication; the hidden state at the last time step is input into the fully connected layer and then passed through softmax to obtain the probability of each risk level.

4. The ICU monitoring customized voice hierarchical alarm system based on personalized voiceprint as described in claim 1, characterized in that, The personalized speech synthesis module includes: The text generation submodule generates broadcast text from a preset template library based on the risk level and warning information; The voiceprint feature extraction submodule stores or extracts the voiceprint feature vector of the target speaker; The speech synthesis submodule uses a neural network-based text-to-speech (TTS) model, which combines voiceprint feature vectors to synthesize the broadcast text into speech, and adjusts the speech rate and fundamental frequency parameters according to the risk level.

5. The ICU monitoring customized voice hierarchical alarm system based on personalized voiceprint as described in claim 4, characterized in that, The speech synthesis submodule adopts the Tacotron 2 architecture, the vocoder uses WaveGlow, and speaker embedding vectors are introduced in the encoder-decoder.

6. The ICU monitoring customized voice hierarchical alarm system based on personalized voiceprint as described in claim 4, characterized in that, The adjustment of speech rate and fundamental frequency parameters according to risk level includes: in, For speaking speed, Based on basic speaking speed, For the fundamental frequency, Based on the fundamental frequency, , For adjustment coefficients, This is a quantified value for the risk level.

7. The ICU monitoring customized voice hierarchical alarm system based on personalized voiceprint as described in claim 1, characterized in that, The vital signs data include at least one or more of blood oxygen saturation, heart rate, and blood pressure; the warning information includes bed number, patient identification, and abnormal vital signs values.

8. A method for customized voice-based graded alarms for ICU monitoring based on personalized voiceprints, applied to the system described in any one of claims 1 to 7, characterized in that, Includes the following steps: S1: Real-time acquisition of patient vital signs data; S2: Classify the current risk according to predefined rules and / or machine learning models, and output the risk level and corresponding early warning information; S3: Generate broadcast text based on risk level and warning information, and synthesize the text into speech signal using a preset specific human voiceprint model, where different risk levels correspond to different speech synthesis parameters; S4: Broadcast the aforementioned voice signal.

Images