Patient Vital Monitoring using AI

Non-invasive monitoring of blood analyte levels using electrocardiograms (ECGs) and machine learning. The technique involves training a machine learning model using ECGs and corresponding analyte levels from multiple individuals. The model can then predict analyte levels from new ECGs without needing blood samples. The training data is processed to filter out potentially inaccurate analyte measurements that could negatively impact model accuracy. This involves analyzing the ECGs to identify features that correlate with analyte levels, as some subtle changes in the ECG may indicate analyte fluctuations that are not obvious to the human eye. The model is trained using these identified ECG features rather than just the analyte measurements. By leveraging the relationship between ECGs and analyte levels, the model can accurately predict analyte concentrations without invasive blood draws.

Source

Health monitoring using multimodal sensors and external data to accurately diagnose diseases and track symptoms over time. The method involves continuously monitoring health using sensors like microphones, peak flow meters, pulse oximeters, etc. It also collects non-sensory data like medication, diet, location, etc. The sensory and non-sensory data is fed into neural networks to detect symptoms and diseases. The networks adapt periodically based on user data to improve accuracy and reduce false alarms. The method also generates personalized audio summaries for doctors to help diagnose.

Source

Wearable electronic device for monitoring health conditions of individuals like seniors with dementia or those at risk of urinary tract infections. The device has a hybrid wireless communication system that can selectively receive location data from indoor Bluetooth beacons and outdoor GPS. It transmits location over low-power wide area networks. This allows accurate indoor and outdoor location tracking. The device also sensors for environment, activity, and physiology. It analyzes the data using onboard machine learning to classify health conditions. Alerts are sent when the wearer leaves a zone. The device connects to a server for training the machine learning model.

Source

A wearable biosensing device that provides personalized health recommendations and risk assessments directly to patients. The device captures biosensing data like blood flow and oxygen levels using sensors mounted on the skin. This data is analyzed using machine learning models to predict health metrics like potassium levels. The predictions are then presented to patients via a graphical user interface on the device. This allows patients to monitor their health without blood tests and get personalized recommendations based on their biosensing data.

Source

Analyzing electrocardiogram (ECG) data using artificial intelligence to accurately and efficiently process ECG signals with enhanced accuracy. The system involves delineating the ECG waves, grouping beats together, and embedding the beats into a lower dimensional space for classification. This allows identifying hidden P waves, handling noisy signals, and handling multiple beats at once. The AI algorithms are trained on large datasets of ECGs from different patients.

Source

Smart mattress that monitors patient health, helps caregivers, and automatically adjusts patient positioning without human assistance. The mattress has sensors to collect biometric data, pressure sensors to measure patient weight distribution, actuators to deform the mattress, and a control system to coordinate motion. The mattress can detect vital signs, pressure ulcers, and posture issues. It can also identify falls and provide alerts. The mattress can autonomously adjust positioning at intervals to prevent bedsores. The mattress communicates with external devices like wearables and EMR systems.

Source

Multi-sensor health monitoring platform that provides personalized health monitoring using non-invasive, wearable, and point-of-care sensors in a patient's home to complement hospital and clinic testing. The platform leverages AI to predict patient needs, energy requirements, and sensor activation based on physiological and activity features. It schedules sensor recharging and identifies which sensors to activate. The platform collects sensor data, applies AI to generate insights and recommendations, and provides tailored care.

Source

A single neural network model for analyzing asynchronous bio-signals from multiple leads without requiring separate models for each lead combination. The model learns to derive an analysis value by reflecting the correlation between leads regardless of the specific combination acquired. This is done through encoding the lead position in the input features and using self-attention computations to propagate the correlation. The model can accurately predict cardiovascular disease even when not all leads are available, since it learns the overall lead interdependence.

Source

Automatic assessment of patient health using a cloud-connected, AI-powered health machine. The machine interacts with patients to collect initial data, determines risk factors, then asks further questions based on the risks. It processes the combined data using machine learning models to determine assessments, diagnoses, and treatment recommendations. The machine provides a comprehensive, point-of-care system for automated patient evaluation.

Source

A patient health management platform for managing metabolic diseases using continuously collected biosignals. The platform generates personalized treatments for metabolic diseases like diabetes by analyzing a unique combination of continuous biosignals from wearables, lab tests, nutrition, meds, and symptoms. It uses machine learning models to understand how biosignals impact metabolic health, generates time series of metabolic states, and recommends treatments based on the patient's profile.

Source

Cost-effective and accurate health monitoring system using wearable sensors and a low-cost processing unit. The system acquires multiple biomedical signals like ECG, PPG, motion, voice, and body temperature using a wearable sensor package. It processes these signals using machine learning models to monitor parameters like arrhythmia, heart rate, blood pressure, respiratory rate, and coughs. The models are trained on a conventional processor and then implemented on a low-cost processing unit. This allows real-time monitoring on commercial devices without needing expensive specialized hardware.

Source

Diagnosing health conditions using patient time series data like ECGs to improve accuracy and timeliness compared to relying solely on derived metrics. The method involves training neural networks on cohorts of patients with the condition versus controls to diagnose the condition based on the raw time series data. It preprocesses the time series by extracting metrics like QT interval, then aggregates network outputs for multiple metrics to generate a diagnosis. This leverages the full time series rather than just discrete metrics.

Source

Assistant robots that can observe signs of health issues, diagnose conditions, and provide medical assistance in homes, workplaces, and healthcare facilities. The robots use sensors to monitor vital signs, movements, and environments, and leverage machine learning models to interpret the data and determine if health dangers or distress is present. If so, the robots can take actions like alerting emergency services, administering first aid, and fetching medication. The robots can also provide regular health checks and treatments.

Source

Configuring input data for a neural network-based model to estimate physiological variables like blood pressure from non-invasively measured bio-signals like PPG. The input data is generated by combining bio-signal data from different times, allowing the model to learn variations in physiological variables over time. This trained model can then estimate current physiological values by inputting the combined historical data and a calibration value.

Source

Non-contact image-based blood pressure measurement using visual intelligence techniques to estimate blood pressure by analyzing video of a person's face. A neural network model is trained using webcam videos to extract features from pulse waves in the videos and predict blood pressure. The model can continuously estimate blood pressure without contact by analyzing videos from a webcam. It leverages visual intelligence to extract features directly from pulse waves instead of using cuff-based measurements.

Source

Predicting ventricular fibrillation (VF) onset in individuals at risk for sudden cardiac arrest using machine learning models trained on heart rate variability (HRV) features extracted from temporal beat activity. The models are trained on HRV from beat activity samples with and without VF events to learn patterns indicative of imminent VF. Inference involves applying the trained models to real-time beat activity to predict VF onset times.

Source

Unobtrusive monitoring system using ultra-wideband (UWB) radar to detect and classify patient motion patterns as indicative of undesirable events like falls, seizures, or sleep disturbances. The system learns normal behavior using machine learning on sensor data to adapt to individual patients. It generates alarms and insights based on deviations from normal patterns. The radar is placed in the patient's environment like a ceiling or nightstand. The system can also integrate other sensors like accelerometers, microphones, and vital sign monitors.

Source

Facilitating real-time estimation of physiological conditions like heart rate, blood pressure, and respiration rate during live video consultations using on-device AI. The system involves a client computer with a camera, storing a trained neural network to estimate physiological conditions from video, obtaining a live video stream during a session, providing video as input to the network to get condition estimates, and sharing those estimates with the remote client. This enables accurate, real-time monitoring and assessment of vital signs during telemedicine consultations using just the client's own camera.

Source

Determining vital signs and physiological states of patients using smart devices like smartphones with cameras, and initiating actions based on abnormal findings. The method involves capturing images of body parts using the device's camera, extracting vital signs like heart rate and oxygen saturation from the images using image processing, and using machine learning models to determine physiological states like hypertension or hypoxia. If the state is outside a threshold, it triggers actions like initiating a video call or making a phone call to a medical professional.

Source

Combining mobile cardiac telemetry (MCT) and Holter monitoring techniques to provide enhanced cardiac event analysis using machine learning. The method involves simultaneously analyzing ECG data for both real-time MCT monitoring and retrospective Holter monitoring. It uses a machine learning model to automatically detect cardiac events during MCT, determine event severity, and send notifications. For Holter, it uses the same model to classify event rhythms. By combining the studies, it provides more detailed analysis than MCT alone and retrospective rhythm classification like Holter.

Source

Using an artificial intelligence model trained on electrocardiogram (ECG) data to estimate cardiac ejection fraction, which is a key indicator of heart health. The model takes an ECG as input and outputs an estimated ejection fraction for the patient. This provides a non-invasive and more widely available alternative to traditional methods like echocardiograms for screening and monitoring ejection fraction. The AI model can be run on consumer-grade ECG devices to enable more frequent and accessible ejection fraction screening compared to specialized equipment.

Source

Using machine learning models to derive health indicators from user-generated content and biometric data to identify root causes of physical and mental health conditions. The models analyze aggregated user data to generate priority lists, predictions, diagnoses, and recommendations for individuals. It involves combining user-generated content like surveys with biometric data from wearables, sensors, and medical devices to identify indicators of health issues. The models associate those indicators with biometrics to predict root causes.

Source

Fall detection using body-worn sensors like wrist-worn devices with improved accuracy compared to existing wrist-worn fall detectors. The method involves a hybrid approach combining decision tree ensemble (DTE) and gated recurrent unit (GRU) machine learning models for fall detection. It first uses a rule-based algorithm to identify potential falls from accelerometer data. These potential falls are then fed into the DTE model to classify as true falls or false alarms. False alarms are further evaluated by the GRU model. By combining results from both models, it improves fall detection accuracy compared to using just DTE or GRU alone.

Source

Using predictive modeling based on heart pump system data and patient factors to forecast patient outcome and track condition. The model determines a heart health index indicative of patient recovery likelihood. Features like pump performance metrics and clinical parameters are used. By quantifying overall heart health, it allows objective ranking/triage of patients to prioritize care. The index can also be used to alert clinicians of declining health.

Source

Detecting multiple abnormal beats in electrocardiogram (ECG) signals using a deep learning method that considers global features of consecutive beats. The method involves extracting global features for each beat, combining them with position information, applying an attention mechanism to weigh the features, and using bidirectional LSTM to learn the weighted feature sequence. This allows detecting and classifying multiple abnormal beats based on the learned patterns of global features across consecutive beats.

Source

A non-invasive system for predicting blood glucose levels using photoplethysmography (PPG) signals. The system involves training a neural network using PPG signals and reference glucose levels from subjects not undergoing medical treatments that can affect cardiovascular health. This allows the network to accurately predict glucose levels for subjects without medical treatments, as opposed to a universal model that includes treated subjects. The network can also use HbA1c levels along with PPG to further improve predictions. Excluding treated subjects improves accuracy as medical treatments can alter PPG signals.

Source

Detecting abnormal blood flow during ultrasound scans using machine learning to identify pathologies like fetal distress. The method involves acquiring Doppler measurements during a scan, evaluating the flow profile to detect abnormalities, and displaying the profile with an overlaid normal reference. This highlights areas of deviation. The system uses trained models to classify vessel types and adjust the reference profile to match the scan. It can also automatically position the Doppler cursor.

Source

User monitoring system using wearable devices to collect physiological and environmental data, analyze it to predict changes in physiological parameters, and alert users about potential issues. The system involves a wearable sensor that collects physiological data from the user's body. It analyzes the data to determine correlations between different physiological parameters. It then predicts changes in a targeted physiological parameter based on correlations with other physiological and environmental data. The system can alert users if predicted changes indicate potential issues.

Source

System to support medical interventions like balloon angioplasty by predicting adverse events during the procedure using machine learning. The system takes in images of the catheter in the vessel and pressure readings. It uses a trained neural network to predict events like balloon rupture or twisting based on the input data. This allows real-time monitoring and visualization of potential issues during the intervention. It can also report and store the predictions.

Source

Classifying PPG signals using scatter diagrams and AI to analyze heart rhythm without requiring long-term monitoring. The method involves extracting inter-beat intervals from PPG signals to generate scatter diagrams, which are then fed into a convolutional neural network for classification. This allows confirming heart rhythm types from PPG signals without needing extensive signal acquisition duration.

Source

Using a camera to remotely measure respiration rate in challenging conditions like low light, by applying convolutional neural networks (CNNs) to video of the torso. The CNNs process pairs of consecutive frames to generate vector fields indicating respiratory movement. Segmentation masks can be used to filter out non-torso vectors. This allows accurate respiration rate estimation from video without requiring physical contact.

Source

Remotely monitoring cardiac condition of patients wearing wearable heart monitors using machine learning analysis to identify worsening cardiac conditions. The method involves applying a two-tier machine learning engine to physiological data like ECG and cardio-vibrational signals collected by wearable devices. The first tier classifies the raw signals. The second tier combines the results with clinical info to determine cardiac condition. Comparing to historical measurements identifies patients with worsening condition.

Source

Real-time monitoring and analysis of healthcare data from medical devices in patient homes to detect issues and trigger emergency procedures. The system involves building a virtual model of the physical device and training patients to use it. Sensor data from the device and patient is transmitted to a cloud service. AI algorithms analyze the data to score the device and patient performance. If the scores exceed a threshold, an emergency procedure is triggered. This allows remote monitoring and intervention to address issues as they arise in real-time.

Source

Monitoring and detecting activities of people using non-wearable sensors like radar, microphones, and cameras to provide accurate and reliable activity detection without relying on user compliance or forgetfulness. The technique involves a sensor array with passive sensors like radar, microphones, and cameras in a home. The sensors capture signals like backscattered radar, audio, and video that are processed using AI to extract information like vital signs, physical activity, and sleep patterns. The processed data is analyzed over time to create baselines and confidence levels. Alerts are triggered when deviations exceed thresholds.

Source

Early detection and monitoring of cytokine release syndrome (CRS) fevers in CAR-T cell therapy patients using continuous temperature monitoring and machine learning. The method involves collecting temperature data from a continuous monitor at high frequency, feeding it to a trained ML system to predict future temperatures, analyzing the predictions for fever indications, and alerting patients/clinicians when fever is detected. This provides earlier fever detection compared to infrequent temperature checks to enable timely CRS treatment.

Source

Extracting discriminating features from patient data for early sepsis prediction in ICU patients. The features are extracted from vital signs, lab tests, and demographics using a sliding time window. The features are normalized and ranked using a relevance method. A classifier trained on the features predicts sepsis risk. It aims to provide a reliable, timely sepsis prediction tool using available patient data.

Source

Device and method for accurately calculating stroke volume from an arterial blood pressure using AI. The technique involves pre-training a stroke volume calculation model using filtered data, then transfer learning it using more filtered data to generate a customized model specific to a patient. This allows higher accuracy compared to conventional methods while being less invasive. The customized model is then used to calculate stroke volume for the patient's arterial pressure.

Source

Early prediction of fatal symptoms like cardiac arrest to improve patient survival rates by using machine learning to classify vital signs data. The method involves converting individualized vital signs to analysis data, generating prediction results for fatal symptom occurrence within a time window, and providing the results to external entities. It uses a modified GAN to train a classification model with augmented data to overcome class imbalance. The GAN generates similar data for each label, allowing label-specific augmentation. The classification model is then trained using both real and augmented data.

Source

Accurately monitoring vital signs like blood pressure using wearable devices without cuffs or mechanical sensors. The technique involves using optical sensors to detect changes in blood volume under the skin. Machine learning models trained on reference device measurements are used to estimate blood pressure from the optical signals. This allows continuous, non-invasive, and accurate vital sign monitoring using compact wearables.

Source

Predicting hemodynamic parameters like blood flow velocity and pressure inside a patient's aorta using an artificial intelligence (AI) system. The AI system takes as input the shape of the aorta and a flow profile around it. It uses a recurrent neural network (RNN) like a Long Short Term Memory (LSTM) network to iteratively predict the hemodynamic parameters at each point along the aorta. The AI system learns to accurately predict the internal flow dynamics based on the vessel shape and external flow data, providing a faster and less invasive alternative to CFD simulations for assessing aortic conditions.

Source

A computer-aided method for detecting life-threatening conditions like sepsis, cardiac arrest, and respiratory failure using biosignals like ECG, EEG, and PPG. The method involves training convolutional neural networks (CNNs) on multi-dimensional power spectral densities (PSDs) from multiple biosignals. It uses ensembles of CNNs or Bayesian CNNs to learn relationships between PSDs from different organs. This allows detecting life-threatening conditions earlier than traditional methods by leveraging concurrent spectral features across organs. The CNNs provide probability estimates for the conditions, which can be displayed to clinicians to indicate increased risk. The method also filters and displays the condition risk over time to provide context.

Source

Efficiently using machine learning in medical devices like implantable cardioverter defibrillators (ICDs) for arrhythmia detection by conditionally applying machine learning models based on feature delineation results. The method involves using feature extraction to identify arrhythmias and checking if the features meet criteria for machine learning verification. If so, machine learning is applied to confirm the arrhythmia. This allows leveraging the lower power and computational efficiency of feature extraction for most arrhythmia detection, while using machine learning selectively for verification.

Source

Discriminating true positive electrocardiogram (ECG) episodes with abnormal rhythms from false positive episodes with normal rhythms using machine learning to reduce false alerts from implantable cardiac monitors. The method involves extracting features from ECG segments and subsegments, feeding them to a machine learning algorithm trained on labeled episodes, and using the output scores to classify episodes as true positives or false positives. This automated analysis replaces manual review of false positives by physicians.

Source

Automatically annotating physiological waveforms using machine learning to provide more advanced and flexible annotation capabilities for analyzing and interpreting patient data. The method involves analyzing a physiological waveform based on a model, automatically generating annotations, displaying them to the user, allowing modifications, saving the modified annotations, and training the model using the modified annotations. This allows more accurate and detailed annotations compared to manual methods.

Source

Identifying apnea and hypopnea during sleep using ultra-wideband (UWB) radar to improve accuracy and robustness compared to relying solely on labeled data. The method involves capturing user movement during sleep using a UWB sensor, then applying Fourier transforms to extract frequency components. Neural networks are used to extract features from the time domain and frequency domain movement data. These features are then fed into another neural network to identify sleep apnea and hypopnea. This multi-step feature extraction approach from both time and frequency domains allows more robust sleep analysis compared to just raw movement data.

Source

Using deep learning to predict future cardiac conditions like atrial fibrillation, aortic stenosis, cardiac amyloidosis, and stroke from electrocardiogram (ECG) data. The method involves training a neural network model on ECG data and supplementary patient information like age and sex. The model generates a risk score indicating likelihood of the condition within a certain time period after the ECG. This score can be used to inform patient care by identifying those at elevated risk and potentially taking preventive actions.

Source

Using a smartphone camera and AI processing to accurately measure blood oxygen levels and capillary refill time (CRT) non-invasively from fingernail images. It involves capturing video of squeezing and releasing a fingernail to analyze changes in color and blood flow. A CNN-based system extracts features, concatenates metadata, and correlates non-spatial data to determine oxygenation and CRT. This provides quantitative metrics for monitoring blood circulation and oxygenation levels using just a smartphone camera.

Source

Predicting cardiac arrhythmias like ventricular fibrillation, atrial fibrillation, and ventricular tachycardia before they occur using multi-modal patient monitoring data. The method involves training a multi-arm deep learning model on extracted local and contextual features from ECG and vital sign data. The model analyzes real-time patient monitoring data to predict imminent arrhythmias and identify contributing factors. This allows proactive intervention before arrhythmias occur. The model architecture has parallel subnetworks for ECG beats, context, and spectrograms.

Source

Method for determining the amount of oxygen required by a user/patient with respiratory problems using AI and wearable devices. It involves collecting oxygen saturation and heartbeat data during a cardiorespiratory function test. A user's behavioral model is computed from this data. Then during daily activities, the model is adjusted based on ongoing oxygen saturation and heartbeat measurements. This customized model is used to estimate the user's oxygen needs at any time.

Source

Using AI and millimeter wave radar to monitor patient anatomical positions for pressure ulcer prevention. A neural network is trained using radar point clouds representing patient positions and durations, along with attributes like temperature and humidity. The neural network predicts pressure ulcer likelihood. Backpropagation adjusts the weights to increase prediction accuracy. This allows real-time monitoring of multiple patients using a single radar, identifying areas prone to ulcers for repositioning.

Source