Secure real-time management of respiratory ventilators and ai-based patient care recommendations

The handheld device with AI-based decision support addresses compliance and communication issues in intensive care units by integrating with ventilators and EMRs, enhancing guideline adherence and reducing errors through real-time data management and personalized recommendations.

WO2026236225A1PCT designated stage Publication Date: 2026-11-12CONVERGENT ENGINEERING INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/026705
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-05
Filing Date
2026-05-05
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Clinicians face challenges in maintaining compliance with best practice protocols, ensuring effective communication during shift changes, and managing high patient-to-clinician ratios, particularly in intensive care units, leading to potential medical errors and suboptimal patient care.

Method used

A handheld device that integrates with respiratory ventilators and EMRs, providing real-time data management and AI-based decision support, including standardized data visualization, personalized recommendations, and compliance monitoring, to enhance clinician efficiency and situational awareness.

Benefits of technology

Improves adherence to clinical guidelines, enhances communication during shift changes, and optimizes ventilator settings, reducing preventable errors and improving patient safety and care efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026026705_12112026_PF_FP_ABST
    Figure US2026026705_12112026_PF_FP_ABST
Patent Text Reader

Abstract

A system and method for managing respiratory ventilators are disclosed. One or more ventilators are communicatively coupled to a computing system configured to receive ventilator-related and patient-related data, including physiologic, demographic, diagnostic, and laboratory information. The computing system converts the received data from a native format into a standardized data format and presents the data in substantially real time through a graphical user interface (GUI). Updated patient data received from the ventilators automatically refreshes the GUI and is used to generate system prompts for an artificial intelligence (AI) system. The AI system produces recommendations for improving ventilated patient care, including ventilator setting adjustments, and selected recommendations may be automatically applied to the ventilator. In some embodiments, the AI system utilizes curated clinical practice guidelines, hospital protocols, medical literature, knowledge graphs, decision trees, and flow charts to generate deterministic and verifiable recommendations.
Need to check novelty before this filing date? Find Prior Art

Description

PATENTSECURE REAL-TIME MANAGEMENT OF RESPIRATORY VENTILATORS ANDAI-BASED PATIENT CARE RECOMMENDATIONSCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority from and is related to United States Provisional Application No.63 / 800,097, entitled “Esophageal Pressure Clinical Decision Support System,” filed May 5, 2026, which is hereby incorporated into the present application by reference in its entirety.BACKGROUNDField of the invention

[0002] The present invention generally relates to physiologic monitoring and decision support for patients. More particularly, the present invention relates to providing clinical decision support, remote patient monitoring, and remote access to the monitoring and decision support, especially for patients who are breathing with the support of a ventilator, supplemental oxygen, or other support devices.Background of the invention

[0003] The present invention relates to numerous papers disclosed in an information disclosure statement filed herewith. To avoid long citations to these papers, these papers are referenced by superscript numbers, e.g., [1,2], and the teaching of each paper is hereby incorporated by reference in its entirety.

[0004] Respiratory Therapists (RTs) and other clinicians deliver life-supporting strategies for patients receiving mechanical ventilation. RTs typically care for multiple patients simultaneously and face several challenges in providing high-quality patient care, including maintaining compliance with best practice protocols, efficient communication during shift changes, and situational awareness with high patient -to-clinician ratios. Although this background often focuses on RTs and respiratory care, all of the issues discussed also apply to other clinicians, especially in the intensive care unit (ICU) where data and monitoring are key to good patient care. In particular, the system can help improve care when used by nurses, physicians, physician assistants, and other specialists.

[0005] RTs and other clinicians are expected to apply best clinical practices as determined by evidence -based clinical practice guidelines (CPGs). Deviations from these CPGs may diminish patient safety and negatively impact patient outcomes. Despite widespread availability and distribution of CPGs, adherence is often low, and clinician adoption is slow. An excellent example is the lung-protective ventilation (LPV) guidelines that recommend the use of low tidal volume (VT) and limited inspiratory plateau pressure (Pplt). These strategies have proven to lower ventilator-induced lung injury (VILI) and mortality in patients with acute respiratory distress syndrome (ARDS).Attorney Docket No. 9224-P0007-WO 1Studies have shown that more than 15 years later, low VT use remains poor, jeopardizing safety and putting patients at increased risk. A 2016 publication showed that even in a hospital that participated in the LPV studies, only 27% of patients received low VT within 48 hours of ARDS onset.SUMMARY

[0006] A handheld device for clinicians, including RTs, that obtains data directly from the Electronic Medical Records (EMR), mechanical ventilators and / or bedside physiologic monitors, with efficient data entry for events, notes, and therapies, can provide decision support customized for RTs or other clinicians that will improve patient care, increase compliance with best practice protocols, enhance safety, improve hand-off communication, and increase clinician efficiency and situational awareness.

[0007] The invention relates to systems and methods for secure, real-time management of respiratory ventilators using artificial intelligence (Al)-based clinical decision support. The disclosed technology integrates one or more respiratory ventilators with a computing system capable of receiving ventilator-related and patient-related data from multiple sources, including ventilator settings, alarms, physiologic measurements, demographic information, diagnoses, and laboratory results. The system converts incoming data from disparate formats into a standardized format, enabling continuous monitoring and centralized visualization of patient status through a graphical user interface (GUI). The GUI is automatically updated in substantially real time as new patient information is received, thereby allowing clinicians to observe current ventilator conditions and patient responses dynamically.

[0008] The invention further includes the creation of Al -generated clinical recommendations for ventilated patients. Updated patient data is used to generate a system prompt for an Al system, which analyzes the standardized data and produces recommendations intended to improve patient care and optimize ventilator settings. In certain embodiments, the recommendations may be automatically applied to the ventilator or displayed to clinicians for review. The Al system may rely on curated source materials such as clinical practice guidelines, hospital protocols, and clinical literature to improve recommendation quality and consistency. The GUI may also be automatically updated with ventilator setting recommendations when the recommendations fall within predefined safety or confidence thresholds.

[0009] In additional embodiments, the invention enhances reliability and explainability of Al recommendations by combining curated medical source materials with deterministic structures such as knowledge graphs, decision trees, and flow charts. These structures may be automatically extracted from clinical guidelines or medical literature using large language models (LLMs). The invention also contemplates the use of natural language summaries of clinical guidelines and protocols to improve interpretability by the Al system and facilitate clinician verification. The summaries may include condition-specific rules and instructions for determining patient-specific conditions such as COPD, ARDS, obesity, or other clinical factors relevant to ventilator management. The Al system may additionally estimate patient conditions using combinations of ventilator data, laboratory data, diagnosis information, and demographic information to generate individualized ventilator recommendations.

[0010] The invention also provides advanced prompt-engineering and recommendation formatting features intended to improve usability for clinicians. These features include instructions directing the Al system to present concise recommendations first, followed by detailed rationale and supporting references. Additional prompt features may include automated calculation of patient-specific parameters, structured markup formatting for dynamic data Attorney Docket No. 9224-P0007-WO 2insertion, and instructions to disregard stale or outdated patient data. Such capabilities improve readability, clinical workflow efficiency, and reliability of Al -assisted recommendations in critical care environments.

[0011] Further aspects of the invention relate to compliance monitoring and protocol enforcement. The system compares updated patient-related data against clinical practice guidelines, hospital protocols, and clinical literature to identify noncompliant ventilator settings or patient conditions. The GUI may visually designate data that falls outside established standards and provide recommendations for returning the patient to compliance. In certain embodiments, Al-generated recommendations are integrated into this compliance framework to support clinician decision-making. The invention also supports automated generation of customizable shift reports that summarize updated patient information and clinician inputs for shift-change communications. Clinicians may revise or supplement these reports with rationale for ventilator changes, treatment responses, completed procedures, anticipated procedures, and recommendations for subsequent care teams.

[0012] Additional embodiments address management of time -sensitive clinical data. Updated patient-related information may be stored in a standardized repository together with timestamps identifying when data was collected from ventilators or related systems. The system calculates time-since parameters indicating the age of each data element and uses those parameters to improve recommendation reliability. For example, the GUI may visually indicate stale data, disable recommendations based on outdated measurements, or trigger additional recommendations when clinical conditions persist beyond predefined durations. The invention also contemplates temporal alignment of delayed data sources, such as laboratory results, with real-time physiologic data to improve analytical accuracy. Multiscale averaging and trend analysis may further be applied to patient data to identify evolving clinical patterns and support increasingly sophisticated ventilator management recommendations over time.BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The accompanying figures wherein reference numerals refer to identical or functionally similar elements throughout the separate views, and which together with the detailed description below are incorporated in and form part of the specification, serve to further illustrate various embodiments and to explain various principles and advantages all in accordance with the present invention, in which:

[0014] FIG. 1 is an image of a dashboard display showing several patients in several rooms, according to one aspect of the present invention;

[0015] FIG. 2 is an image of a similar dashboard designed for phones with smaller, narrower screens, according to one aspect of the present invention;

[0016] FIG. 3 is an image of a higher-level dashboard that might be used by clinician managers who are overseeing entire units or groups of units, according to one aspect of the present invention;

[0017] FIG. 4 is an image of a detailed patient information page, typically displayed when clicking on a patient or room on one of the dashboards, according to one aspect of the present invention;

[0018] FIG. 5 is an image of the data collection boxes that collect data from medical devices and pass them to the computing resource, according to one aspect of the present invention;Attorney Docket No. 9224-P0007-WO 3

[0019] FIG. 6 is an image of a shift report display on the user interface, according to one aspect of the present invention;

[0020] FIG. 7 illustrates an example of a notes entry in the system, according to one aspect of the present invention;

[0021] FIG. 8 depicts an embodiment of the system’s network topology, according to one aspect of the present invention; and

[0022] FIG. 9 illustrates preliminary results from a study of the system, indicating that the system increases compliance with guidances, according to one aspect of the present invention.

[0023] FIG. 10 depicts a block diagram of an embodiment of the system to create a query or prompt for an Al system to create optimized recommendations for improving patient care

[0024] FIG. 11 depicts a block diagram of an embodiment of the system in which format conversion to a standard language occurs before transmission to the server.

[0025] FIG. 12 depicts a block diagram of an embodiment of the system in which an Al system with a curated library of documents, CPGs, protocols, and clinical literature is used to generate a recommendation that is checked for reasonableness before display.

[0026] FIG. 13 depicts a block diagram of an embodiment of the system in which the curated library of documents from FIG. 12 also includes extracted information from papers or CPGs, typically in the form of decision trees or graphs.

[0027] FIG. 14 depicts a block diagram of an embodiment of the system where CPGs, literature, and protocols are used to create a rule-based system to determine compliance of the data with the library and then provide guidance to improve compliance with the library

[0028] FIG. 15 depicts a block diagram of an embodiment of the system where the server is monitoring settings of the medical device or ventilator continuously and automatically determines when changes are made to the settings. When changes to the settings are detected, the system asks the clinician to justify the change to the medical device, and this justification is stored in the database for future use, such as in a shift report.

[0029] FIG. 16 depicts a block diagram of an embodiment of the system in which the server uses data collected during a shift to automatically fill in a shift report template and shows this shift report to the clinicians to improve shift report communication.

[0030] FIG. 17 depicts a block diagram of an embodiment of the system in which timestamps are included with all the data, and this is used to time-align sensitive data and calculate a “time-since” parameter that allows for showing the age of the data.DETAILED DESCRIPTION

[0031] As required, detailed embodiments are disclosed herein; however, it is to be understood that the disclosed embodiments are merely examples and that the systems and methods described below can be embodied in various Attorney Docket No. 9224-P0007-WO 4forms. Therefore, specific structural and functional details disclosed herein are not to be interpreted as limiting, but merely as a basis for the claims and as a representative basis for teaching one skilled in the art to variously employ the present subject matter in virtually any appropriately detailed structure and function. Further, the terms and phrases used herein are not intended to be limiting, but rather, to provide an understandable description of the concepts.

[0032] The description of the present invention has been presented for purposes of illustration and description, but is not intended to be exhaustive or limited to the invention in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the invention. The embodiment was chosen and described in order to best explain the principles of the invention and the practical application, and to enable others of ordinary skill in the art to understand the invention for various embodiments with various modifications as are suited to the particular use contemplated.

[0033] Overview

[0034] A challenge faced by clinicians is the consistent transfer and receipt of critical patient care data during shift changes. Poor “hand-off’ communication has been shown to be a major factor in preventable medical errors, directly affecting patient safety. While it sounds simple, a high-quality hand-off is complex and rare in practice. A 2016 study estimated that communication failures were responsible for 30% of malpractice claims in U.S. hospitals, resulting in 1,744 deaths and $1.7B in malpractice costs over five years. While the use of a computerized hand-off method can result in improved information transfer during the transition of some care providers. There is a dearth of such tools available for RTs and other bedside care staff.

[0035] Other compromises to safety in patient care are associated with understaffing or the ineffective use of the clinician’s time. Understaffing is a chronic issue in US hospitals and diminishes the clinician’s ability to provide appropriate patient care and interventions. The COVID- 19 pandemic further stressed staffing, heightening the need for tools to improve efficiency and the rapid adoption of new standards and care methods.

[0036] Compliance with Best Practices

[0037] Every year, the National Institutes of Health (NIH) invests roughly $32 billion in medical research, and in 2012, the total estimated expenditure on biomedical research in the US was $120 billion. More than 30,000 clinical trials are registered with clinicaltrials.gov each year. Unfortunately, the resources expended on this research usually require years, if not decades, to effectively reach the public. For example, tremendous effort and resources are spent to help determine the best clinical practices. In practice, however, implementing best practices into routine clinical care is complicated and often takes more than a decade. There are a variety of barriers to guideline adherence, including lack of awareness, lack of familiarity, lack of agreement, lack of outcome expectancy, and inertia of previous practice. Many of these barriers could be overcome by enhancing communication and consolidating key patient information in an RT friendly format.

[0038] Poor Communication during Shift Changes

[0039] Health care provider discontinuity is an unfortunate but necessary reality of longitudinal hospital care. This discontinuity creates opportunities for errors when clinical information is not accurately transferred between providers. A study of malpractice claims involving Emergency Room (ER) physicians found that communicationAttorney Docket No. 9224-P0007-WO 5failures among providers are a major cause of preventable error. In an observational study of 10 hospitals, recipients reported that 37% of hand-offs were unsuccessful. Another multicenter study demonstrated that implementing a standard hand-off bundle named I-PASS could reduce preventable adverse events by 30%, without increasing the time of transition of care. Several key actions suggested by The Joint Commission to improve the quality of transition of care include standardizing content between sender and receiver, including standardizing tools and methods (e.g., checklists), face-to-face hand-off in a quiet environment, use of electronic health record capabilities and other technologies (apps and patient portals) to enhance hand-offs, and process improvement through monitoring the success of interventions.

[0040] In a 2014 hospital survey of 405,281 staff members, 53% endorsed the statement that “important patient care information is often lost during shift changes,” and 47% endorsed the statement that “shift changes are problematic for patients in this hospital”.

[0041] High Patient-to-Clinician Ratios

[0042] RTs, like other clinicians, are often burdened by heavy workloads, including the need for time-consuming documentation that prevents them from spending adequate time at each patient’s bedside. This highly demanding workflow often makes it difficult to obtain appropriate situational awareness. The North Carolina Respiratory Care Board states that understaffing is common in respiratory care (RC) services and places patients at risk of diagnostic or therapeutic errors, thereby increasing hospitals' liability. In a 2016 Minnesota survey of RTs, the greatest source of dissatisfaction amongst RTs was feeling overburdened (in their words: short-staffed, stretched too thin, with a high patient-to-RT ratio, and / or the workload increasing without more staff). In a California Society of Respiratory Care survey of 130 RC Managers, 30% indicated they do not have adequate staff, and 21 % indicated significant and chronic understaffing; 30% of these directors reported patient safety issues resulting from the understaffing. Staffing problems will likely continue since the Bureau of Labor Statistics indicates that the need for new RTs will grow 23% over the next 10 years, and 30,000 more RTs will be needed.

[0043] Decision support tools and the EMR

[0044] When designing a system to support clinicians and improve patient care and safety, the natural inclination is to build Clinical Decision Support (CDS) tools into the EMR or to use data directly obtained from the EMR. However, few systems have been able to leverage this abundant data to provide patient safety improvements. [1] While the EMR plays a crucial role in medical billing and documenting the patient’s medical history, a lack of complete, actionable information and delays in manual data entry are significant barriers to developing real-time CDS tools. Furthermore, in many places, the combination of privacy laws and the IT infrastructure created by the EMR has created additional barriers, such as restrictive data access and security policies that appear to have made the data even less accessible than in the past. [2,3] Also, most hospital IT systems are a patchwork of old and new systems that can be significantly different even among different floors of the same unit. As such, healthcare data is rarely standardized, often fragmented, or stored in incompatible formats. [3] Additionally, due to data storage requirements, most EMRs do not store or display waveform data that can be used to extract useful data for intelligent decision support. Lastly, because most healthcare systems are heterogeneous [4] and difficult to maintain, there is often significant inertia in hospital IT departments against modifying or allowing access to EMR systems.Attorney Docket No. 9224-P0007-WO 6

[0045] Existing ICU Data Tools: Over the last 20 years, many ICU data visualization tools have been developed, but none thus far have made a significant impact on the market. Tools such as AWARE [5], take data from the EMR and provide dashboards and remote monitoring capability. A recent review article stated that Randomized Controlled Trials (RCTs) of physiologic and laboratory monitoring displays did not show improvement in primary outcomes, despite positive results in simulated settings. [6] We believe that dashboards and checklists can provide process improvement, but what is needed is decision support to improve outcomes.

[0046] Al in Mechanical Ventilation: Al has been used for many years in Mechanical Ventilation (MV) to predict conditions such as ARDS and estimate important parameters such as Work of Breathing. More recently, reinforcement learning (RL) approaches have been proposed to dynamically optimize MV, specifically oxygenation and tidal volume. Lee et al. provide an approach that extracts decision trees from the system, calling it an “interpretable RL system”. [7] Although exciting, these approaches are limited in scope, face regulatory challenges, and are validated using historical data and off-policy evaluation, which is known to produce high variance results (making it difficult to test interventions with historical data). [7] Lastly, these methods only learn from data, whereas our system is derived from CPGs and can provide physiologic rationale, both of which are key advantages.

[0047] Although MV is sometimes thought to be mainly protocol-driven and dependent on only a few measured parameters, rule-based decision support algorithms can capture only a small part of the overall complexity of the MV optimization process. There are many different MV protocols based on patient condition, disease status, and stage (stabilization, weaning). Selecting the right protocol and MV settings should be done by assessing the patient data in the context of the patient's physiology and maladies. In fact, Tobin states that “no area of critical care demands a greater understanding of physiological principles than ventilator management” and that use of protocols alone can lead to complications and risk of death. As such, although compliance to protocols and CPGs improves patient care, they must be used in the right situations and based on sound physiologic assessment, not blindly followed. An Al system that blends the corpus of available CPGs and protocols, assesses the patient physiology, and provides recommendations with physiologic rationale is a significant improvement compared to rule-based protocol following techniques.

[0048] Decision Support Tools and LLMs: A large language model (LLM) is an advanced artificial intelligence system based on deep learning and transformer neural networks, trained on massive datasets to understand, generate, and predict human language text with contextual awareness and high accuracy. Artificial Intelligence (Al) and LLM approaches to improving healthcare are expanding and improving dramatically. LLM examples include personalized medical guidance, clinical decision support, identifying urgent medical situations, and communicating with patients with empathy. ChatGPT has achieved surprisingly good results on medical licensing exams in a variety of fields and countries, and has simplified documentation by writing discharge summaries and operative reports. A significant benefit of incorporating LLM recommendations into treatment strategies lies in its untapped capacity for crafting individualized medical guidance. Although LLMs can provide excellent responses to some clinical questions, current LLMs are not optimal for clinical decision support. A significant challenge for the use of LLMs is their imperfect accuracy and / or tendency to produce hallucinations - outputs that are factually incorrect or irrelevant. Providing correct, hallucination-free results is especially important in healthcare. Secondly, references / citations from LLMs are often unreliable. LLM responses are also stochastic, meaning that the same question can result in different answers each time.Attorney Docket No. 9224-P0007-WO 7

[0049] Aspects of the claimed invention can improve patient care by:• Providing remote patient and device monitoring, increasing situational awareness, and providing decision support for the clinician,• Improving communication during shift changes using a standardized hand-off,• Providing nudges, indicators, and advice to improve compliance with best-practice protocols and CPGs,• Allowing for continuous improvements, such as rapidly tracking changes in CPGs, prioritization of tasks and alarms, and advice for improving patient safety.

[0050] The system is a novel and marketable product because:• New, powerful, and inexpensive loT devices dramatically reduce system costs and development time.• It is device-agnostic and can be used with any medical device that exports data.• It adds advanced features, such as CDS, to existing devices without requiring replacement.• The user interface can be operated on any device with a web browser (mobile, PC, etc.).• The invention can transfer full-resolution waveform data for enhanced feature extraction and CDS, or it can process the waveform data in situ and transmit only the derived parameters to the system.• Initial releases will likely not require Food and Drug Administration (FDA) premarket clearance, accelerating the first release of the product.• The platform allows for a continually expanding set of clinical decision support.• The platform can be extended to provide CDS to a variety of specialties.

[0051] Additional features include:• Protocol support and therapy reminders - prioritizing and reminding the clinician when therapies and procedures are due will improve efficiency and patient care (e.g., “Spontaneous breathing trial is due”).• Better management of device alarms - the Emergency Care Research Institute (ECRI) states ventilator alarms are a top 10 Health IT Hazard, and ventilator disconnects are in the top five list of hazardous alarms. Real-time patient monitoring enables several improvements in alarm management, including alarm prioritization, predictive alarms, and the reduction of nuisance alarms.Attorney Docket No. 9224-P0007-WO 8• Detection of ARDS, Chronic Obstructive Pulmonary Disease (COPD), and other conditions that affect care - an important barrier to implementation of the LPV recommendations is the delayed recognition of ARDS. Using existing standardized definitions of ARDS, COPD, and other respiratory diseases, we can warn clinicians that these conditions exist, enabling earlier intervention to prevent worsening.• Advice on optimizing device settings and therapies based on device data (such as real-time ventilator and physiologic monitor data), and extracted waveform features.• Advice on inspiratory muscle conditioning / training - often completely ignored, maintaining and restoring inspiratory muscle strength is critical to liberating the patient from the ventilator. Before the development of ventilators, respiratory muscles were never idle, and research has shown that respiratory muscles atrophy very easily, even after a few hours of rest during surgery. As such, it is critically important to carefully monitor patient muscle activity and provide extubation support to avoid difficult-to-wean patients - a major contributor to high-cost ICU stays.• Machine intelligence to optimize shift reporting by automatically determining key changes in patients - to assist clinicians in more efficient shift change reporting, key changes in the patient status could be automatically detected by the system to ensure that they are communicated to the oncoming clinician.

[0052] Because the invention has the potential to provide these features without replacing expensive medical devices or upgrading the EMR, it represents an innovative method of assessing and enhancing patient care and safety.

[0053] For clarity, we will first describe the usage model for an embodiment of the invention. For each patient (indicated by room or bed number), large boxes display important parameters such as LPV parameters ( Vr / kg and Pplt), along with alarms and scheduled activities. Device settings and patient data are also shown. For each patient, there are buttons to document events or activities, as well as a shift report button. The “notes / events” button allows the clinician to update activities and procedures done throughout the day (e.g., transport, suctioning, BronchoDilator (BD) treatments, etc.) and the “Reports” button formats the patient information, device settings, and activities / procedures into a structured shift report.

[0054] More specifically, FIG. 1 is an image of a dashboard 101 display showing several patients in several rooms denoted by room number 102. The dashboard display provides a quick overview of each patient and guidance or protocol deviations color-coded in 3 different levels (red, yellow, grey) 104. The dashboard contains a snapshot of important information for each patient in a minimalist fashion to maximize the number of patients that can be seen simultaneously. The dashboard allows for a small number of parameters, called watches, to be shown that may be clinically relevant to the clinician 105. These watches can be manually selected by the clinician 107. In a preferred embodiment, these watches are triggered automatically by the system. For example, when a ventilator change is recognized that may affect oxygenation, the SpO2 measurement will be displayed automatically in the dashboard. This watch may also be automatically turned off after a certain amount of time or after the patient is determined to be stable. Similarly, watches can be automatically turned on when the patient starts a spontaneous breathing trial, an awakening trial, or has certain treatments or procedures done. The dashboard 101 contains a brief summary of theAttorney Docket No. 9224-P0007-WO 9demographic and diagnosis data 103 for the patient for quick access by the clinician. In one embodiment, the demographic data and diagnosis data 103 displayed is chosen automatically from the plethora of EMR data based on the type of patient and the type of clinician using the system. For example, respiratory therapists are focused on the respiratory system and, to a lesser extent, the cardiac system of the patient. The EMR contains a large number of conditions the patient may have, and each clinician-type can have a list of pertinent conditions that will be displayed. In one embodiment, the system shows which patients are connected to a ventilator 106 and are often the highest acuity patients.

[0055] FIG. 2 is an image of a similar dashboard designed for phones with smaller, narrower screens 201. The dashboard shows the room numbers assigned to the clinician, basic patient information, and guidances / protocol deviations. In a preferred embodiment, the room or bed numbers are used as the primary linking data between data from different devices. The use of a patient identifier for this purpose creates data security and privacy issues. In this more compact view, only the room / bed number, basic patient information, diagnosis, and guidance alerts are shown. Clicking the patient opens the patient detail view, which provides all relevant patient information and data. Due to their high acuity, ventilator-based patients can be sorted to the top of the list if desired 202.

[0056] FIG. 3 is an image of a higher-level dashboard that might be used by clinician managers who are overseeing entire units or groups of units. In this dashboard 301, only room numbers and the number of guidance / protocol deviations are shown. This gives a high-level view of the units and how many patients need attention. This view shows at a high-level which patients require the most attention and which units have the highest acuity levels.

[0057] FIG. 4 is an image of a detailed patient information page, typically displayed when clicking on a patient or room on one of the dashboards 401. The patient information page shows all the relevant data for a single patient (room / bed). It shows the patient information 408, patient diagnosis 409, guidances or protocol deviations 402, measurements from medical devices 405, settings from medical devices 407, blood gas, or similar lab data from the EMR 406. It also has a tab to create watches on the dashboard 410, a tab for events that are recorded (including changes in device settings or guidance status) 403, a tab for entering data 404, a tab for viewing or editing shift reports [not shown], and a tab for Al recommendations [not shown],

[0058] FIG. 6 is an image of a shift report display on the user interface 601, according to one aspect of the present invention. It is usually formatted in a standard shift report format such as SBAR or IPASS. The SBAR format, for example, includes Situation, Background, Assessment, and Recommendations sections. The system automatically takes data collected from the devices and clinicians throughout the shift, along with EMR data if available, and automatically formats it to support high-quality shift reports. Hand-typed notes can be used to fill in data not automatically obtained from the EMR 603. Notes entered by the clinician are automatically placed in the correct section of the report 604. Lab data is entered automatically and timestamped 605. Timestamps are important for all data because older data is less relevant than newer data, among other reasons.

[0059] A history of medications delivered, treatments given, and other things that occur throughout the day are included 606. A list of events, including automatic detection of device settings changes and explanations of those changes 607. Guidances that are triggered on or off during the shift are also recorded as events. Recommendations are typically hand-entered. Shift reports can be imported from previous shifts to simplify some data entry 609. In this case, only a subset of relevant data is copied, leaving out shift-dependent information like treatments given, etc. ShiftAttorney Docket No. 9224-P0007-WO 10reports can be saved, and at the end of the shift, in one embodiment, after the shift report, are finalized and saved for future reference 610. Note boxes and other information can be manually entered by the clinician throughout the day or just before shift change.

[0060] FIG. 7 shows an example of notes entry in the system 701. There are tabs with different categories of notes typical of the specialty of interest (in this case, respiratory therapy notes). In a respiratory care embodiment, categories are treatments 702, assessments 704, suctioning 705, ventilation 706, and complications 707. Each category may have several note types, and when clicked on, it asks additional questions or provides a text box for free-form data entry. The preprogrammed note types and note-detailed questions allow quick, painless data entry of standard notes kept by the clinician while minimizing required typing. Selecting assessment 704 will bring up additional questions with preprogrammed possible entries , such as delivery method 708 and medication name, as part of a category Chest Physiotherapy (CPT) as shown. The notes entry also allows for the entry of complications or similar entries.

[0061] In a typical facility, some devices, such as ventilators and infusion pumps, are mobile and wheeled into the room before use, while other devices, like the physiologic monitor, are typically mounted in the room. To ease the determination of which device is attached to which patient, a preferred embodiment will label each device (ventilator, monitor, etc.) with a short numeric code with QR code (like a bar code) that uniquely identifies the device. When a new patient enters the unit, the clinician will enter or scan the mobile devices to link them to the room number or bed number in the hospital. The fixed devices can be programmed to each room once or often have the room number entered into the device. The system can read this room number from the output of the device. If the system is connected to the EMR, the patient information and demographics can be automatically linked to the room / bed number. If not connected to the EMR, the clinician may have to manually input key information (such as height, weight, sex, diagnosis, etc.) into the system. If a patient changes rooms or new equipment is used, the process is quickly and easily repeated.

[0062] As an alternative approach, one data collection box is associated with each room / bed number. When a mobile device (e.g., ventilator) is used in the room, it is either wirelessly or via a communication wire (e.g., serial cable, USB, ethemet cable) attached to the data collection box, thus linking the mobile device to the room. The fixed devices in the room can be permanently connected to the data collection box. If the connection to the data collection box is done wirelessly, the clinician will need to verify either (a) that the connection made to the device is correct (that the device connection is correct for this patient), or (b) manually enter the device ID to link it to the room number and patient. A key feature of this approach is to associate the patient and devices with a room number or bed number rather than directly with a patient identifier, which is typically protected health information and more difficult to use.

[0063] FIG. 5 is an image of the data collection boxes 501 that collect data from medical devices and pass them to the computing device that implements the service, according to one aspect of the present invention. The computing device can be in the cloud, a local server, or an edge device that implements the system locally. The boxes have a serial port (502), an Ethemet connector for wired internet (503), a Wi-Fi component for wireless internet [internal], Bluetooth [internal] for wireless connections to medical devices, and power [not shown],

[0064] It is important to note that different medical devices (e.g., ventilators or physiologic monitors) will typically have proprietary (e.g., different) communication protocols that must be properly interpreted by the data collection box. There are a plethora of medical device communication ports and protocols. The physical layer of mostAttorney Docket No. 9224-P0007-WO 11medical devices includes RS-232 serial ports, USB, Ethernet, Wi-Fi, Bluetooth, and other interfaces. The data is often transmitted in manufacturer-defined protocols that determine the order and exact format of the data transmitted. On other devices, the data is transmitted via industry standards to simplify interaction. These standards include HL-7, ICD-10, SNOMED CT, LOINC, FHIR, and others. Aspects of the invention include the ability to read data from a variety of physical layers and to interface with many different protocols. In a preferred embodiment, the data collection box automatically determines which device is connected without user intervention. In one embodiment, the data collection box is notified by hardware when a device is connected to a port and begins sending queries to the new device at multiple baud rates until it receives a message consistent with the response expected from a known device, indicating it has found the correct device protocol. In a second embodiment, when there is only a small number of different devices to communicate with, the data collection box has multiple interface ports preprogrammed for different devices. Data interpretation and translation to a standard format can occur at the data collection box, the server, or an intermediate device connected to multiple data collection devices. It is important to understand, though, that the data collection box must have sufficient knowledge of the devices it needs to connect to properly configure the communication port and ensure it is receiving valid data.

[0065] In an embodiment, the interface to the data collection box, when not wireless, includes a physical-layer conversion to a standard wired format to interface with a variety of devices with different physical layers. In one embodiment, this physical layer conversion is low-cost and passive, such as a wired dongle that converts a DB9 connector to an RJ11 connector. This inexpensive converter can be applied to all devices and remains with that device to simplify connection to the data collection box. In other cases, this physical layer conversion might include a USB-serial device or a serial-USB device.

[0066] Once the proper device driver is found, the system must interpret the data, alarms, device settings, and other information received from the medical device. Since different manufacturers and models use different communication protocols and different terminology (non-standardized format), it is important to translate data from each model into a common or standardized format for storage and display by the system. In one embodiment, this translation is implemented using a table lookup that associates parameter numbers or parameter names with a standard list of parameters (similar for settings, alarms, and other data). For example, one manufacturer may use the terminology VTe as parameter 95 for expired tidal volume, and another may use TidalVolumeExpired as parameter 7 in block 9 for expired tidal volume. Some devices may not have a value for expired tidal volume. Using this common terminology throughout the system greatly simplifies data storage and use within the system, as well as data interpretation by the clinician. In one embodiment, the data collection box performs this translation by using custom protocol parsers (drivers) for each medical device supported by the system. In this method, the system has virtualized protocol-interpretation engines (drivers) for each available device, which can be started one by one to determine which device is connected to the box. In another embodiment, the data is sent to the server, where it is translated and utilized. In either method, the data will be sent and stored securely, typically by encrypting it with a secure communication protocol and a secure storage format. The data received by the server is stored for use in shift reports, data analysis, time alignment with other data, and analytics and quality-improvement metrics. This common terminology can be custom-designed for the system or based on standardized clinical terminologies such as HL7 or similar formats. HL7 or similar formats.

[0067] Since data arrives from different sources at different times, it is important to add timestamps to the data and store them with the data. For example, blood gases are obtained infrequently (e.g., daily) and may not remain Attorney Docket No. 9224-P0007-WO 12valid for more than a few hours after collection. As such, displaying timestamps or greying values that are more than a few minutes or hours old (depending on the data) is important to ensure the clinician knows the age of the data they are analyzing.

[0068] The system has a built-in function called “time since” that calculates the time difference between “now” and the time the data were measured (for example, from a device or the time the blood gas was drawn for lab data). This “time since” parameter has a variety of useful purposes, including the ability to change the color of the data as it ages, thus gracefully de-emphasizing older data. The time-since parameter also simplifies disabling guidance or recommendations that depend on data older than a certain number of seconds or minutes. Some alerts or guidance are also time-sensitive, such as automatically detected changes in the ventilator or new blood gases or x-rays that should be reviewed - ideally, these guidances should not remain on for more than an horn or so. In addition, the time-since parameter enables even more complex multi-layer guidance to better match written protocols based on when the guidance has been active. For example, often times, a protocol or CPG will say that an action should be taken if one or more values are outside of range for more than x minutes, which is easily obtained using one second -level guidance that is triggered on one or more simpler guidances indicating parameters are out of range.

[0069] In addition, timestamps are necessary to link data from different sources that happen roughly at the same time. Again, using a blood gas as an example, it may be reported several hours after the blood was drawn. To compare the blood gas data with the real-time data, a timestamp (in this case, multiple timestamps - one for the time drawn, one for the time reported) is required. For example, PaCO2-ETCO2 is a common parameter that indicates the expected difference between an infrequently measured blood gas value (PaCO2) and a real-time sensor value (ETCO2). But it is important that these numbers be synchronized with the time of blood gas draw to ensure the values are correct, because ETCO2 can change dramatically with changes in ventilator settings or patient-ventilator interactions.

[0070] For automatic guidance or alerts, time-based values are often used instead of instantaneous values. For example, for highly variable parameters such as SpO2, a 1 -minute or 10-minute average SpO2 may be better than an instantaneous SpO2 value for clinical decision support. These multi-scale averages are also useful in determining short-term and long-term trends. For example, a long-term upward trend can be determined by subtracting the 1 -hour average from the 24-hour average and checking whether the result is significantly above zero. Similarly, a short-term negative trend can be determined by subtracting the 5 -minute average from the 1-hour average and checking whether the result is significantly less than zero.

[0071] Since the MRN is considered PHI, if it is used, it will be stored via a cryptographic hash function that will create a 1 : 1 mapping to the patient without storing the MRN. In this way, linkages between devices, patients, room numbers, units, and RTs can be made dynamically without storing PHI.

[0072] Shift Reports

[0073] Shift report communication is critical in improving patient care and decreasing medical errors. However, shift report communication is known to be poor, and improving it is a focus of many quality improvement programs. However, good shift report communication requires either good documentation and / or more time than many clinicians are willing to spend, especially when they are about to leave the hospital. To help improve shift report, communication without increasing time required by the clinician, the system automatically imports data into an electronic (can also Attorney Docket No. 9224-P0007-WO 13be printed) shift report format using data automatically obtained by the system, including device-measured data, device settings data, device alarm data, EMR data such as demographics, and other data. Since data can be obtained from various sources and device manufacturers, it is first translated into a common language or standard format. In addition, notes taken throughout the day in the system can be easily added to the shift report and pushed back into the EMR (if available), reducing documentation requirements. When used electronically in the system, the shift report created by the off-going clinician can be dynamically reviewed simultaneously by both the off-going and on-coming clinicians, either by sharing a single large screen or by synchronizing the shift report screens on two different displays. Notes can be dynamically entered and viewed by either the on-coming or off-going clinician as they discuss the patients being transferred, using the dynamic shift report format.

[0074] Since shift report formats will vary from unit to unit and hospital to hospital, the electronic shift report formats will be easily modified to support a wide variety of clinicians and units. In one embodiment, the customizable shift reports will use a markup language to organize text and the standardized protocol format for the variables obtained by the system. For example, the markup language includes bracketed variable names and standard text with font commands. For example: [Bold] Tidal Volume: <VTe>[ / Bold]. In this way, customization can be easily achieved and new quality goals can be appropriately included and documented as needed.

[0075] Since quality shift reports summarize the state and care a patient is receiving, storing the shift reports and making them available to clinicians provides a meaningful history for determining changes in the patient's status and care. This data can be used for analytics, patient care analysis, quality metrics, and outcome analysis as well. Allowing an on-coming clinician to import the previous shift report into their shift report simplifies completing a shift report. In a preferred embodiment, a predefined list of parameters / settings / notes / etc. are automatically imported into a shift report. Some fields should not be imported; they should be filled out dynamically during the new shift by the system or the clinician. For example, treatments given should not be copied from the previous shift report.

[0076] Throughout the shift, the clinician will document procedures and activities that are completed and enter any new blood gas results, diagnoses, or issues. Because the tool is specifically designed for clinicians, this data can be quickly and efficiently entered with little required typing except for freeform notes. For example, when an endotracheal tube (ET-Tube) suctioning is documented, the system pops up a window asking for the color, amount, and thickness of secretions with radio buttons. In most cases, these pop-ups require less than 10 seconds to fill in.

[0077] At the end of the shift, the tool greatly simplifies the shift report. Since the clinician has incrementally documented important information, there is no preparation time required for shift report since all data is already stored in the system. Both the sending and receiving clinicians will have access to the system and will go through the structured hand-off (shift report) together. Notes can be added as required by either clinician and both will have access to the data and reports from previous shifts.

[0078] Although physicians often have mobile tools for recording orders and notes, RT s and other clinicians often use a paper task list and handwritten notes to record their daily activities. The invention will be an efficient, easy to use electronic replacement for this paper-based system.

[0079] Guidances And AdviceAttorney Docket No. 9224-P0007-WO 14

[0080] A key feature of the invention is the ability to offer advice based on clinical practice guidelines, hospital protocols, or other documents that are used to guide care. Clinicians are often slow to adopt new CPGs or protocols, and electronic support has been shown to increase compliance. Increased compliance with protocols and guidelines not only helps standardize care but also has been shown to improve outcomes. Although most CPGs and protocols are described in text, many are data-driven and can often be expressed as equations or similar combinations of data converted into the system’s standard language parameters. The key difference between this protocol support and standard alarms is that alarms are typically just a threshold on a single parameter, whereas CPG / protocol “guidances” are often a combination of several parameters and / or mathematical derivations from several parameters. For example, for the Lung Protective Ventilation CPG, the exhaled tidal volume should be divided by Ideal Body Weight (a function of patient height), and the resulting parameter should be between 4 and 8 ml / kg. Also, plateau pressure, which is usually measured during an end-inspiratory pause, can be estimated dynamically using exhalation time constants (or other methods), yielding a dynamic plateau pressure that remains less than 30 cm H2O. This parameter can also be checked against peak inspiratory pressure ( PIPO, which has to be greater than or equal to plateau pressure, to doublecheck the accuracy of the estimated plateau pressure. Likewise, some CPGs only apply to patients with certain conditions, for example, ARDS, which is readily available in either notes entered by the clinician or diseases documented in the EMR or calculations of oxygen versus FIO2, which can be imported directly when available. These combinations of parameters and patient conditions provide more helpful information and guidance for clinicians and analytics / quality measures for managers as they relate directly to hospital protocols and CPGs - unlike standard alarms. In addition to detecting when the patient is outside the protocols or CPGs, the system may use the protocols or CPGs to recommend changes in device settings or other patient care techniques (such as treatments, suctioning, etc.) to improve outcomes and bring the patient back into compliance. Because the guidelines and recommendations are directly derived from existing protocols and CPGs, the preferred embodiment does not create “new” decision support; rather, it electronically implements known rules and protocols, which will simplify regulatory approval for the device.

[0081] Since the patient's diagnosis is important for determining the appropriate source material and recommendations, and the medical record may not be available or contain up-to-date information, an automated system to detect conditions is a valuable addition to the system. Some conditions are determined easily; for example, obesity can be determined by assessing BMI, which is calculated from height, weight, and sex. Another example is the Berlin criteria for ARDS which uses PaO2 / FIO2 ratio and imaging and timing to determine ARDS and severity. A surrogate for PaO2 / FIO2 ratio is the SpO2 / FIO2 ratio that can be used for more real-time detection of ARDS. Other conditions may be more complicated and require Al or machine learning tools.

[0082] An example guidance system in the invention is the Oxygenation guidance -the latest guidelines indicate that maintaining saturation (SpO2) between 88% and 96% during use of supplemental oxygen improves outcomes. This produces a warning when the saturation or PaO2 is outside this range, leading to a recommendation to increase or reduce PEEP and / or FIO2. In addition, the ARDSNet recommended a table of PEEP and FIO2 settings that help the clinician choose the appropriate pairing of PEEP and FIO2. This is easily programmed into the system and the guidance can help the clinician maintain the settings in compliance with the ARDSNet table.

[0083] To help reduce situations where the guidance system may recommend unsafe changes, a preferred embodiment of the system is a method for checking the sanity or safety of the recommendation after it is generated but before it is displayed. In a simple embodiment, this safety check is implemented by comparing the Attorney Docket No. 9224-P0007-WO 15recommendation against a set of settable thresholds that limit the system's ability to recommend settings. For example, a setting outside of a hospital protocol or CPG set bound would either modify or prevent the GUI output of the recommendation. In a more complex embodiment, the bounds are set by the patient's diagnosis, status, or demographics, and may also include higher-level maximum and minimum bounds. In a third embodiment, the system is compared with physiologic model outputs or previous patient conditions to predict or estimate the safety of those settings on a more personalized basis.

[0084] Aspects of the present invention provide a real-time clinical intelligence system that integrates multimodal real-time data streams, including one or more of electronic medical records (EMR), ventilator parameters, physiologic monitors, infusion pumps, and / or laboratory results, to generate predictive analytics and Al-driven forecasts of critical events in intensive care unit (ICU) patients. By processing high-frequency, time-series data through machine learning models trained on historical patient outcomes, the system identifies early warning patterns indicative of impending clinical deterioration, such as sepsis, acute respiratory distress syndrome (ARDS), hemodynamic instability, or ventilator-induced lung injury. The models can employ deep learning architectures, including recurrent neural networks (RNNs) and transformer-based algorithms, or other models to dynamically adapt to patient-specific physiology and contextualize data relative to population-level trends. Predictive outputs are translated into actionable clinical alerts (e.g., “48-hour sepsis risk probability: 82%”) and intervention recommendations (e.g., “Increase PEEP to 12 cmEEO; initiate fluid resuscitation”), enabling proactive care. Because the system dynamically obtains data and outcome data, it can optimize itself and learn over time to improve performance as new data is acquired. This learning can use any number of learning models, but the preferred embodiment optimizes itself via continuous reinforcement learning, refining predictions based on real-world clinician responses and patient outcomes. This innovation can reduce unplanned ICU transfers, mitigate preventable adverse events, and shorten ventilator dependency by aligning therapeutic decisions with individualized, anticipatory analytics.

[0085] A key request of clinicians is determining when the settings on a medical device like a ventilator are changed. In some cases, nurses, physicians, physician assistants, and therapists may all change the device's settings, and the clinician may forget to document the change. For this reason, guidance is provided that indicates that the device settings have been changed. This can be determined by constantly monitoring the device settings (which are often exported along with device data) and determining when a key parameter has been changed. Some parameters that are not important enough to trigger a “settings change guidance” can still be changed; they are first converted to the common device format and then checked against the table of settings that meet the required level. At this time, guidance will be shown (and recorded) on the dashboard and the patient's detailed display. In one embodiment, the guidance can be automatically dismissed when the EMR indicates that the settings change has been documented. In that case, the users of the system can be notified of who made the change to the device. In an alternative embodiment, the system asks the user to explain the rationale for the ventilator change, perhaps using a list of standard rationales. The rationale and the settings change and associated timestamps are stored on the server and used for analytics and shift reporting. If an input communication to the EMR is available, this rationale can be pushed to the EMR to document the device settings change.

[0086] Another feature is analytics and management-quality reporting, including quality -tracking and reporting functions that allow managers to track individual, shift, and unit-level performance. Using the data to improve adherence to standards and patient care and being able to track improvements over time would provide value to hospital management.Attorney Docket No. 9224-P0007-WO 16

[0087] Management quality reporting is implemented with a separate login for managers. This account enables graphical and printable graphs and tabular reports (exportable to Excel) to track performance by RT management. The reports will be adjustable for different customers, and the data and analytics are stored in a searchable format, such as an SQL database. This allows for the quick and easy customization of management, quality, and analytics reports.

[0088] IT requirements for installing the system typically included: IT / Cyber security (preventing unauthorized access to computers, networks, and data) was the number one priority. Most facilities have an IT security department, and a lengthy process is required to pass the IT security evaluation. As suggested by several hospital data security managers, we plan to implement a secure and robust medical device design process and have it certified by a 3rdparty certification entity. HITRUST is the most used certification in the healthcare environment, but a full HITRUST certification can only be achieved after the product has been fully commercialized. All data transmission and storage in the system will be encrypted. For physical security, the Data Aggregator will be located in a hospital-controlled, secure room and will only allow secure login using a public / private key scheme. Several hospitals require that we install our system on a blade server that will be installed in the hospital's IT rack. Other hospitals require that we operate in a virtual machine on an existing network. Both types of installations are achievable with the present invention.

[0089] FIG. 8 shows an embodiment of the system’s network topology 801. The medical devices 803 (both fixed and mobile) are connected to the network “nodes” 802 via wired or wireless methods, and then transmit data on the local network (either wired or wireless) to the cloud-based software system 804, which organizes, stores 805, and displays the data to the clinician. The EMR feed 806, when available, resides on a separate node that may run on a hospital server, where the data is translated and PHI is removed before it is forwarded to the cloud-based, local, or edge-based system. In one embodiment, the system links all data via room number to avoid PHI. In another embodiment, the servers and storage are located inside the hospital to avoid moving patient data outside the control of the hospital system. The system can operate roughly identically whether hosted in the cloud or hosted locally, as the users and nodes will transmit and receive data identically, regardless of the location of the servers operating the core system software.

[0090] The system can operate on several platforms (tablets, laptops, Workstation on Wheels (WOW)) and can quickly and easily customize the GUI to meet the needs of each clinical unit. Moving to a Web Server / Browser architecture simplifies the system's design and supports improved data security features. The data aggregator will include an HTTP server Application Programming Interface (API) that enables almost universal client-side implementations (any system with a web browser).

[0091] We have implemented clinical studies with the high-level goals to show that (a) adhering to CPG guidelines for ventilator-dependent patients is easier with the system, (b) the use of RT Assistant improves shift report communication quality, and (c) RTs find the system easy to use and helpful in caring for their patients. The study included evaluating RTs in two study phases: a Pre-intervention Phase (PP, aka control group) and the Interventional phase (IP). In both groups, we will collect data from physiologic monitors and ventilators (including ventilator settings and alarms, airway waveforms, breathing pattern data, saturation (SpO2), end-tidal partial pressure of carbon dioxide (PetCO2), and hemodynamic data) from adults on ventilators. In the PP, RTs were treating patients in the normal manner without using the system, while in the IP, RTs will treat a similar group of patients with the use of the system.Attorney Docket No. 9224-P0007-WO 17At shift change, hand-off reports for the PP and IP will be monitored unobtrusively (observation-only) by highly experienced, trained respiratory therapists using a checklist of items expected to be transmitted.FIG. 9 shows preliminary results from a study of the system, indicating that it significantly increases compliance with the guidances or guidelines.

[0092] Pre-Intervention Phase: In the pre-intervention phase, the data collection system was installed in the ICU and will operate for approximately 1 month to collect baseline / control data. Study staff constantly monitored the system to ensure that high-quality data were collected reliably.

[0093] Intervention Phase: In the Intervention Phase, the data collection system remained active and the RT Assistant user interface software was installed. Before starting the Intervention Phase, the study coordinator and RT trained the unit RTs on the hospital-specific version of RT Assistant and on the importance of quality shift reports, LPV guidelines, and oxygenation management guidelines.

[0094] Two major methods of evaluating the system were implemented under IRB approval in the Surgical ICU, Medical ICU, and Neuro ICU at the University of Miami. We installed the system and began passively collecting data (no data provided to clinicians) in April 2024. In July of 2024, we began showing the data from the RT Assistant in the clinical office and allowing clinicians to use the system live.

[0095] Guidances in the system are used to notify the clinician when the patient is outside of conventional clinical practice guidelines. The new guidances included in this version of the RT Assistant include:• RR: Respiratory rate should be less than 30• Oxygenation: SpO2 should be greater than 90 and PaO2 should be great than 60• High FIO2: FIO2 should be 40 or less when SpO2 is greater than 95• Pplat: Pplat should be less than 30• PEEP-FIO2: PEEP and FIO2 settings should roughly follow the ARDSNet PEEP-FIO2 table

[0096] The system automatically calculates the percentage of time that these guidances were active (indicating a change in care should be considered). Before the system was available to clinicians (-5000 ventilator-hours across 80 patients), the percentage of times the guidances were active was significantly higher than after the system was available (-4000 ventilator hours across 57 patients). For example, more than 70% reductions were found in warnings related to Respiratory Rate, Plateau Pressure, and PEEP-FIO2 settings. The table below shows the results:Before Afterguidance Available Available DifferenceResp Rate >30 15.3% 3.8% -75%SpO2<=90 1.8% 0.9% -50%Attorney Docket No. 9224-P0007-WO 18FIO2>40 when SpO2>95 25.3% 15.8% -37%Pplat and PIP>30 5.6% 1.5% -73%PEEP / FIO2 table 22.1% 6.4% -71%

[0097] After using the system, we asked RTs and RT managers to evaluate the RT Assistant for usability and effectiveness. Results were excellent and are shown in the table below (l=definitely, 5 = No):average Question scoreCan RT Assistant improve situational awareness? 1.0Can RT Assistant help during your care of ventilated patients? 1.0Did RT Assistant improve the ability to maintain CPGs 1.1Did RT Assistant improve awareness of LPV and oxygenation guidelines? 1.0Can RT Assistant improve patient safety and improve outcomes? 1.2Was the dashboard showing all patients useful? 1.2Was the patient detail screen useful? 1.2

[0098] Network Device Architecture

[0099] FIG 10 shows a method for managing respiratory ventilators that involves communicatively coupling multiple ventilators 1001 or medical devices 1002 and optionally EMR data 1003 to at least one server using a secure communication protocol 1004, such that each ventilator provides ventilator data from a patient being ventilated, called ventilator-related data. Ventilator-related data includes one or more of ventilator settings, ventilator measured parameters, ventilator alarms, or combinations thereof. Ventilator-related data is augmented with one or more of data from other medical devices 1002 and the EMR data 1003 and laboratory data to create a larger set of data called patient-related data that includes all data available for a patient who is being ventilated. The data is stored on the server 1005 in a database 1007 in a standardized format. Remote users 1012 and local users 1013 are provided access to view the patient-related data in real time through a graphical user interface (GUI) 1011. At least one of the ventilators or other sources of patient-related data may provide updated data in a non-standardized format that is dependent on its model. The server converts this non-standardized data into the standardized format 1006, stores the standardized data in a database 1007, and automatically updates the GUI 1011 in response to storing the new data. The system identifies real-time physiological and demographic data associated with the standardized updated data and uses at least part of that data to generate a system prompt 1008 to an artificial intelligence (Al) system 1009,Attorney Docket No. 9224-P0007-WO 19which then produces updated ventilator setting recommendations 1010. The system prompt includes both physiological and demographic patient data.

[0100] FIG. 11 shows that the format conversion 1101, which converts one or more patient-related data such as ventilator, medical device, EMR, and laboratory data 1102 from non-standard formats to a standardized format, may occur outside the server 1103, potentially in the system that reads the data from the medical device before sending it in a secure format to the server.

[0101] FIG. 12 shows that the Al system may further compare the standardized data to a curated library of source materials 1201, such as clinical practice guidelines (CPGs), hospital protocols, and clinical literature. The system may include a sanity check 1202 that determines whether the recommendation is reasonable, for example, by comparing suggested settings against a set of predefined thresholds or common bounds for parameters. If deemed reasonable, the GUI 1203 is automatically updated with Al-generated ventilator setting recommendations. In FIG 13, the Al system may employ a hybrid model to generate a query or prompt 1302 that combines a curated library 1301 with tools such as knowledge graphs, decision trees, and flowcharts 1301 to generate deterministic, verifiable recommendations. These tools may be automatically extracted from CPGs, hospital protocols, or medical literature using large language models (LLMs) as part of the Al system. Clinical experts may verify the source materials.

[0102] In embodiments where at least two ventilators and / or medical devices from different manufacturers are used 1303, each may produce non-standardized data that must be converted 1304. A method for generating personalized ventilator recommendations includes the Al system receiving patient-related data such as real-time physiological and ventilator data, accessing a curated library of guidelines and protocols, analyzing the data using a hybrid retrieval-augmented generation (RAG) system with an LLM 1302, and generating personalized ventilator setting recommendations.

[0103] FIG 14 shows an embodiment in which, instead of using an Al system, the library of CPGs, literature, and protocols (library documents) 1401 is used to create rules 1402 that determine whether the live data from devices 1403, converted to a standard data format 1404, is compliant with the library documents 1401. These rules are used to create guidance 1405 to improve compliance with the library documents and to potentially provide recommendations for improving compliance with the library. In this embodiment, the rules 1402 can be automatically extracted from CPGs, literature, and protocols using an Al system, or manually extracted by the clinician via an interface that allows entry of complex multi-parameter rules. Both the content of the curated library and the rules used by the system can be easily modified by the hospital or by individual units within the hospital to allow fine-grained customization. The rules are used to measure adherence to CPGs and protocols, and statistical analysis is performed to determine whether adherence is improving.

[0104] Another aspect of the invention involves comparing existing clinical protocols against a curated library of the latest clinical literature, with the Al system generating a report that highlights inconsistencies and recommends updates to align the protocols with current knowledge. Additional methods focus on GUI-driven features such as identifying patient-related data that do not comply with CPGs or protocols, visually flagging this data, and offering corrective suggestions.[00f05] FfG f5 shows another embodiment that automatically monitors patient-related data such as from the ventilators and / or medical devices 1501 after it has been converted to a standard format 1502, and detects when Attorney Docket No. 9224-P0007-WO 20settings on the ventilators or medical devices have been changed 1503 by comparing the new settings to the old settings in the database 1504. When a setting change is detected, the system requests that the user 1505 provide a rationale for the change to the device. This rationale is saved in the database 1504 for future use, including for use in the shift report. The data can also be pushed back to the EMR 1506 if available.

[0106] FIG 16 shows further features, including automatically generating customizable shift reports that incorporate patient-related data such as ventilator and / or device data and EMR data 1601 and user input 1602, and displaying those reports on the GUI 1603 during shift changes. The method also addresses interoperability by managing ventilators or other sources of patient-related data from various manufacturers that transmit nonstandardized data formats. FIG 17 shows that these data streams are standardized and stored with timestamps (1701) indicating when the data was generated. The system calculates a time-since parameter 1702 for each data point. It uses it, along with the current time, to temporally align the patient-related data streams, compensating for differences in sampling rates and transmission delays.

[0107] Visual cues, such as color changes in the GUI, may indicate the age of the data (1703), while the system can disable or enable certain recommendations based on whether the underlying data is within or beyond a set freshness threshold. The frequency of key measurements — such as blood gases or baseline values — can be tracked and analyzed to support quality improvement.

[0108] In addition to time alignment properties, temporal and historical data are often used to identify or track trends. A robust, easily implemented method used by the system applies multi-scale averaging of ventilator and other data to detect trends and long-term changes, supporting more complex recommendations, such as “parameter A is falling rapidly”. For example, a 5-breath average can be compared to a 100-breath average to determine whether a parameter shows a short-term trend, and a 30-minute average can be compared to a 2-hour average to determine whether a parameter shows a long-term trend. Additionally, the sampling rate may automatically increase in response to specific events or ventilator changes.

[0109] Machine Intelligence

[0110] Optimizing patient care and mechanical ventilation (MV) involves continuous adjustments to devices such as mechanical ventilators and their parameters to match patient requirements. Incorporating Al into decision-making can enhance patient management strategies, including mechanical ventilation. In a preferred embodiment, we augment large language models with real-time patient data and a library of CPGs, protocols, and articles (referred to as “source material”) to enable the Al to assist the clinician with patient-specific decision-making while maintaining compliance with CPGs, which has been shown to significantly improve outcomes.

[0111] Several methods can be used to optimize LLMs for tasks such as clinical decision support (CDS), including pre-training, fine-tuning, and prompt engineering. Pre-training uses traditional data sets augmented by additional domain data, such as PubMed, to train a new LLM, but requires massive computing resources and significant manual labor. Fine-tuning updates the training of existing LLMs with domain-specific clinical literature but still requires significant effort and costs, as well as access to high-quality sources not already included in the LLM's training set. In addition, custom models are often surpassed in performance by the next release of the general -purpose LLM. A category of customization called “prompting” uses an unmodified LLM but provides detailed instructions and background information to help it provide better responses. Prompting methods include In-Context Learning or ICL Attorney Docket No. 9224-P0007-WO 21(specific queries that stepwise lead the LLM to an answer), Chain-of-Thought or CoT (producing a chain of prompts that break the task down into sequential steps), Prompt Tuning (learning the best prompts to produce the result), and Retrieval-Augmented Generation models that enhance the performance by integrating external knowledge into the retrieval process. For small enterprises, combinations of prompt tuning and RAG strategies can achieve strong results, often outperforming the more costly methods.

[0112] The primary advantage of the RAG approach is that it enable the LLM to access a curated external database of information, thereby focusing its search on highly relevant data and clinical literature. The RAG approach effectively reduces the occurrence of hallucinations and reduces variability in the responses. Without enhancements, LLMs tend to provide generic responses that are of limited clinical value. Using patient-specific information and a curated library, RAG models can improve specificity, enhance response detail, and include more recent data in recommendations. The goal of all of these methods is to use patient-specific information with hospital protocols and CPGs to create patient-specific recommendations and decision support.

[0113] Extraction of Clinical Decision Support (CDS) Information for and from LLMs: A key difficulty in CD S use of LLMs is the repeatability and interpretability of LLM-based solutions. To improve CDS use of LLMs, methods such as decision trees or graphs can be used to make results more deterministic and easier to validate. Hybrid models can use a combination of decision trees or graphs and LLM logic, where simple decisions can use the easily validated, deterministic logic from the decision graphs / trees, and more complex decisions can be made by the LLM using the prescribed library and previous training. The hybrid models can thus product more deterministic and verifiable ventilator setting recommendations and clinical decision support. In addition to the use of decision trees, graphs, and other representations of the source library, it has been shown that natural language summaries or abstracts of the source library can dramatically improve the LLM's performance. Many CPGs are quite inefficient at providing carespecific recommendations, often including 20 or more pages of text for a simple recommendation to keep a parameter less than a certain value. As such, a natural language summary of the different documents in the source library can create a concise, easily verified (by clinicians) representation of the source library for easier access by the LLM. A key attribute of this approach is that the natural language summary improves interpretability of the source materials by the Al system and creates a more accurate and reliable recommendation from the Al system. The natural language document can include preconditions such as “if the patient has ARDS” or “if the patient is not obese,” and simple recommendations on how much a given setting may be adjusted based on the values of one or more patient / system parameters. These documents can also add prioritization of the different papers based on patient conditions or hospital / unit preferences. For example, ETCO2 management may be de-emphasized in certain units relative to oxygenation management.

[0114] Despite the recent increase in the medical use of LLMs, most applications have focused on offline chatbots. The preferred embodiment automatically integrates real-time streaming data from medical devices and the medical record directly into a query submitted to the LLM. The query is optimized to focus and format the response as desired by the clinicians. The query is also optimized to reduce variability and increase the system's testability. Testability and repeatability are crucial in the ability to sufficiently test the system for regulatory approval.

[0115] A key feature of this system is that the LLM is used only to answer a small subset of questions and that it uses real-time data, along with patient demographic and other patient data. This combination of a template query that includes real-time patient and device data greatly reduces testing requirements and regulatory burden compared to aAttorney Docket No. 9224-P0007-WO 22chatbot or similar uses of LLMs. Another feature of the system is that these queries are easily editable by the administrators of the system. Parameters can be switched or modified, limits can be changed, and local hospital rules can be implemented in the query itself.

[0116] In another embodiment, the LLM is augmented with a RAG-system designed to reduce variability in its responses, limit hallucinations, provide verifiable results, and limit computational complexity. The key feature of the RAG-system that provides advantages to clinical decision support is that a curated library of documents, articles, protocols, and CPGs can be created by clinical experts, and the LLM can be asked to limit its reasoning or focus its reasoning on these documents - thus limiting data or source material that is not verified by the clinical experts. This framework with the customized queries greatly improves the repeatability, reliability, and consistency of the resulting recommendations by limiting the background material used by the LLM. When combined with the customized queries that focus the LLM to answer primarily or solely based on the curated library of articles, protocols, and CPGs, the system is even more reliable and repeatable. In addition, this curated library can be optimized for parsing by the system (tokenized, etc.), thus increasing system performance.

[0117] The curated library of source material is a key feature of the system that also allows for hospital -to -hospital or department-to-department customization. For example, the cardiac ICU will likely have more patients with healthy lungs; thus, more focus can be placed on rapid weaning from the ventilator. The medical ICU may have more obesity and COPD. The curated library can include subsections of source material focused on specific disease states, such as obesity, COPD, ARDS, etc. These subsections can be explicitly referenced when the patient's diagnosis (or patient parameters like height and weight) indicates that the patient falls into these categories. The curated library would also typically include hospital guidelines or protocols specific to the local hospital or department.

[0118] Disease and condition-related protocols and CPGs enable personalized patient recommendations. The curated library of source material would include condition-specific material when the disease modifies the best practices for that patient. In some cases, the condition can be as simple as obesity, and in others, it is a disease or condition that affects how the patient will respond to therapy. For example, obese patients typically require more PEEP, and ARDS patients have low compliance and poor oxygen exchange. Both conditions may alter the system's recommendations by including the obesity or ARDS library in the query when appropriate. One might be tempted to add ALL the libraries to create a master library with extensive source material, but this will often make the recommendation system more unreliable, as it has too much material to search through.

[0119] Some papers, protocols, and CPGs can be used to create simple flow charts or decision graphs for clinical decision making. An offline LLM has been created that is used “offline” to extract the key decision-making rules from the source material, and then the resulting graph or flow chart can be used directly in the online decision support to both reduce computational complexity and also improve repeatability and accuracy. The graphs and flow charts also simplify verification and allow for simple modifications that may be necessary from unit to unit or hospital to hospital. This LLM model with RAG and decision graphs is a hybrid RAG model that improves recommendations and further decreases variability and testing complexity. Graphs, flow charts, or simple text-based rules can also be added manually to the curated library to easily include departmental or hospital practice guidelines that are not published. For example, the rule of thumb in most ICUs is that FIO2 is not normally weaned below 40 and almost never reduced below 30%. This is not necessarily consistent with the latest literature, but it is common practice. AsAttorney Docket No. 9224-P0007-WO 23such, these simple graphs, flowcharts, and manual mles can easily augment the curated library to improve performance at each location or for each patient type.

[0120] The invention also includes an off-line tool, in a preferred embodiment containing an LLM, that can automatically or semi-automatically (a) improve / validate written protocols, (b) find conflicts / contradictions in the curated system library, and (c) evaluate how changes in the library affect the system output. Adding local protocols to the curated library enables customization at the unit and hospital levels, which is critical for securing clinician support. However, these modifications can create conflicts or unknowns to the decision support system. As such, having an offline tool to verify that changes to the library have the desired effect is critical to the product's success. Typically, the system is tested with a series of vignettes (clinical scenarios) that represent a wide variety of important patient conditions. In one embodiment, the offline tool will run many vignettes through both the original and modified systems and automatically highlight significantly different recommendations for evaluation by the human reviewers. Since recommendations are textual and may differ in the text but not be substantially different clinically, an LLM or similar logic can directly compare the results to determine whether they are significantly different.

[0121] In another embodiment, a simple product could be used to evaluate a local hospital’s protocols versus state-of-the-art CPGs and the latest literature. The preferred embodiment of this system uses the RAG-LLM model and the latest literature library (curated by experts in the field) to evaluate the protocol. The system also makes suggestions on how to modify the protocol to be more consistent with the latest literature.

[0122] Additionally, the framework can be easily applied to many other specialties such as nurses, PAs, and intensivists, by modifying the library, queries, and user interface.

[0123] The Kaiser Permanente Chief of Al stated that Al should support clinicians in improving care and reducing their workload, and that clinicians are demanding new Al tools. However, the tools must be validated, trustworthy, and designed to fit into the clinician workflow. The invention represents an innovative method for assisting clinicians in improving patient care and safety. It fits naturally in the clinician workflow and addresses the critical issues limiting direct use of Al and LLMs in CDS.

[0124] In developing our Al -based recommendation system, we implemented 3 tasks to build the library and test scenarios.

[0125] Task 1. Collect a library of CPGs, protocols, and respiratory research to support respiratory decisionmaking. We made a numbered list of key clinical decisions and a taxonomy of patient conditions and maladies that affect RT decision-making. Using this list, we addressed each item with protocols, CPGs, textbook chapters, and research articles. When not available online, the key information was extracted manually. Each reference was labeled with keywords and a reference to the numbered list to help ensure completeness. Disease-specific (e.g., COPD, ARDS, obesity) source material was collected and labeled. During development, we created LLM-based tools to find resources, check document consistency, and support rapid development and improvement of hospital protocols.

[0126] Task 2. Extract clinical scenarios for evaluation of the library and system. The preferred testing method for clinical decision support is the use of a benchmark set of clinical scenarios, or “vignettes”. Statistically, random sampling for decision support often misses edge cases and other clinically important situations. As such, we created a benchmark database using “critical case sampling” and “maximum variation sampling”. We created a set of clinicalAttorney Docket No. 9224-P0007-WO 24vignettes that covers the decision types described in the Task 1 taxonomy, including valid combinations of multiple disease conditions, physiologic conditions, ventilation stage, and current ventilator settings. When matching data couldn’t be found, clinicians modified data from similar scenarios (e.g., how would this COPD patient's data differ with sepsis). All scenarios were labeled and clustered to simplify incremental development and testing.

[0127] Task 3. Evaluate the library by manually determining recommendations and rationale for each scenario and ensuring each decision can be justified using the library. For each scenario, our clinicians made high-level recommendations and completed a cross-reference table linking the recommendations to the library entries.

[0128] In developing the Al -based recommendation system, we implemented several tasks.

[0129] Task 1 - Query optimization and baseline performance. Optimized queries are critical to obtaining good results and often contain a list of steps and rules for the LLM to use, such as “reference all source material”, “limit responses to known facts”, “review these documents before answering”, etc. We evaluated the results from different query structures (including stepwise ICL or CoT-like queries - for example, sub-queries for oxygenation, ventilation, and lung protection) and different LLM parameters available in Bedrock, such as Temperature, which regulates LLM “creativeness” and limits hallucinations.

[0130] Further prompt engineering included a series of steps for the LLM to follow to properly evaluate the patient's condition prior to making a recommendation. Examples include calculating BMI to determine if the patient was obese since obese patients require higher PEEP, determining it the patient likely has ARDS by evaluating the PaO2-FIO2 or SpO2-FIO2 ratio, prioritizing certain settings over others (e.g., oxygenation over CO2 removal) when certain conditions exist, etc. By instructing the Al system exactly how to “think” about the current scenario, including evaluating the patient, evaluating the ventilator and physiologic data, and evaluating the needed changes, the recommendations are much more consistent and accurate. In addition, the prompt itself can instruct the Al system to prioritize certain reference documents in the source library, thereby enabling RAG-like performance without strictly using a RAG architecture. Not only can the prompt instruct the Al how to break down the task into manageable pieces, it also instructs the Al system how to report results. In a preferred embodiment, the prompt asks the system to provide a concise summary of recommendations first (since clinicians are often busy), followed by a list of calculations and conditions it performed, followed by a detailed description of why it made the recommendations, followed by a list of references it used. Not all of these steps are required, but structuring the response makes the system much more acceptable to busy clinicians. In another embodiment, the prompt can instruct the Al system to ignore data that is incorrect or out of date, it can instruct the system to use only the material in the curated library, it can instruct the Al system how and when to interpret certain data, and to strictly interpret its learning without extrapolation.

[0131] Task 2 - Create RAG system with library from Aim 1 and optimize the performance. We used a stepwise approach to prototyping the system in AWS Bedrock by embedding documents incrementally using vignette categories (i.e., Aim 1 keywords or XML tags). As we add a new topic / category of documents, we not only validate the vignettes in the new topic but also evaluate whether other previously validated topics are affected. Remember that this is not a chatbot. The system is restricted to specific inputs and queries and is tasked with providing specific advice to clinicians (i.e., improved ventilator settings for patients with X conditions and Y ventilator settings). This stepwise construction and evaluation of the system provides a stable mechanism for adding and testing new material after the system is completed. The system provides a method to test each addition to the library against the entire benchmarkAttorney Docket No. 9224-P0007-WO 25and report which results are “different”. Clinicians will evaluate the differences to ensure consistency across libraries and avoid unintended consequences. Since the recommendations are in plain English and include a physiologic rationale, we created a separate LLM tool to evaluate similarities and differences among recommendations from different system outputs.

[0132] Task 3 - Create a hybrid RAG system for improved performance. As discussed previously, decision trees and knowledge graphs are excellent methods of storing information that are useful for CDS. LLMs support the extraction of decision trees and graphs and their use as input to their decision-making process. For example, we used the LLM to extract a decision tree for PEEP and FiO2 titration from 5 key CPGs / papers, then added it to our LLM query. This led to a more deterministic, predictable recommendation. This method can simplify and codify CDS and can be used to create a deterministic, easily computable first-order response that can be validated, then augmented as needed with LLM queries. The system can augment the RAG system with decision trees extracted from both CPGs and the LLM itself. This has been shown to improve performance, decrease variability, and enable verification.

[0133] The Al system is a first-of-its-kind, real-time decision-support tool for MV. We have shown that using a hybrid RAG model can provide accurate, verifiable, and consistent results, including physiologic justifications. The system will help clinicians comply with clinical practice guidelines and local protocols, significantly improving patient care and clinician efficiency.

[0134] In addition, an Al or LLM model can be trained to determine when parameters received are inconsistent or likely in error. Lab values are notorious for being physically impossible. Heart rate data from the pulse oximeter can be compared with ECG data, if available. Some data combinations are also likely to be in error, such as very low SpO2 and very low ETCO2.

[0135] Al and the Electronic Medical Record: When designing a CDS system, the natural inclination is to build tools into the EMR. However, few systems have been able to leverage this abundant data to deliver real-time improvements in patient safety. The EMR often lacks timely, complete, actionable information, and current EMRs typically lack waveform data and real-time medical device data. Additionally, the combination of privacy laws and complicated hospital infrastructure often creates barriers that limit its direct CDS value. However, advances in ICU data systems may enable future CDS in the EMR.[00f36] A unique feature in a preferred embodiment of the system, when EMR integration is available, involves comparing clinician orders with the patient's current state. The system determines when a new order is received by monitoring the EMR data received and notifies the clinician user. Periodic reminders can be transmitted to the clinician user. The order can be a new treatment, a transfer order, or another type of order. In addition to reminding the clinician user that a new order has arrived, the system can monitor data and user input to verify whether the order has been completed yet. If the order is for a change to a medical device, the system verifies that the change has been implemented and can automatically document it. If the system has an interface to push data into the EMR (or the system is within the EMR), it can automatically document in the EMR that the order has been completed. Similarly, orders for treatments may be detectable in the data stream from the medical device - such as breathing treatments, suctioning, ECG placement, or other monitoring. For orders that are not easily determined from device or EMR data, the user simply clicks the “new order” guidance and can mark it completed, potentially pushing the data “completion” to the EMR. These features represent a drastic improvement in communication between the clinician who places theAttorney Docket No. 9224-P0007-WO 26order and the clinician who completes it - simplifying and accelerating the entire process. In addition, the system can compare the orders with standard practices, local protocols, and clinical practice guidelines and determine if there are contraindications or other problems with the order. In an embodiment, the system indicates this with a warning to the clinician in the “new order” indicator. Another feature is the determination of how long it takes for a clinician to implement an order from the time it is placed. This is a key quality metric that can be automatically calculated, stored, and summarized for clinicians or management.

[0137] Some orders specify protocols, other guidelines, or a series of steps for therapists or clinicians. For example, an order may be written for a therapist to follow the “Adult weaning protocol,” which includes a series of instructions and methods intended to wean the patient from a ventilator gradually . In this case, the order can be used to “enable” a new protocol that the system can use to turn guidance on or off relative to it. It can also be used to include or remove certain protocols or CPGs from the library currently being used by the system to make LLM-based recommendations. All of these methods may help avoid clinical errors and enhance patient safety.

[0138] Other orders may restrict care or prohibit procedures, such as Do Not Resuscitate orders. The system automatically interprets these orders and includes the status in the shift report. In addition, this information is presented on the dashboard and / or patient detail page via color coding or simple indicators to ensure the clinician is aware of these care restrictions. In some cases, these procedures can be automatically determined from device data or EMR and cross-checked against the restrictive orders. If it is determined that the care is being delivered when it should not be, the system notifies the users or management immediately. For example, CPR can be determined from the ECG, pulse oximeter, or other physiologic monitor channels.

[0139] This invention also presents a universal platform for integrating, standardizing, and intelligently managing ventilator alarm signals from diverse device sources, employing multi-layered signal processing to mitigate clinical alert fatigue while enhancing patient safety. The system consolidates heterogeneous alarm outputs from multiple ventilator manufacturers through a proprietary middleware architecture that normalizes disparate alarm taxonomies into standardized clinical significance categories. In a preferred embodiment, Machine learning algorithms continuously analyze temporal alarm patterns alongside patient-specific physiological parameters, electronic health record data, and historical alarm response outcomes to dynamically assign actionability scores. In another embodiment, the platform implements context-aware filtering that suppresses transient, self-resolving alarms (such as momentary high inspiratory pressure events during patient movement) while preserving notification for clinically significant events requiring intervention. A graduated escalation matrix routes prioritized alerts to appropriate care team members via configurable communication pathways, with automated secondary notification workflows triggered when primary responders are unavailable. Another embodiment includes the system using adaptive learning to refine alarm thresholds based on patient-specific baselines, reducing nuisance alarms while maintaining sensitivity to critical events. This comprehensive approach significantly reduces sensory overload among clinicians, shortens response times to actionable alarms, and creates a more conducive healing environment by lowering ambient noise levels in critical care settings.

[0140] The universal ventilator alarm management platform would extend beyond alert handling to serve as a comprehensive quality -metrics calculation system through automated quality -metrics functionality and a regulatory integration framework.Attorney Docket No. 9224-P0007-WO 27

[0141] Automated Quality Metrics Calculation

[0142] In one embodiment, the system leverages its data integration capabilities to automatically calculate critical quality indicators, including but not limited to: (1) VAE Risk Assessments: By analyzing ventilator and patient-related parameter trends (e.g., deteriorating compliance, FiO:> / PEEP increases), the system can implement the validated electronic algorithm for VAE detection that demonstrates excellent sensitivity and specificity compared to human abstractors. This allows for real-time risk stratification rather than retrospective identification. In a second embodiment, the system automatically collects VAE data from the EMR link and stores it along with patient parametric data useful for tracking and future analytic / predictive analysis. (2) Bundle Compliance Tracking: In one embodiment, the platform creates “digital signatures” of ventilator or care bundle elements (thromboembolism prophylaxis, stress ulcer prophylaxis, oral care, daily sedation breaks) with high accuracy, enabling automated compliance measurement without manual chart review. In a second embodiment, the system automatically collects bundle compliance data from the EMR linkage and stores it along with patient parametric data useful for tracking and future analytic / predictive analysis. (3) Complication Frequency Monitoring: The system continuously tracks ventilator-associated complications using standardized definitions, automatically calculating rates using the AHRQ formula: [(# of events) / (Total vent days)] x 1000.

[0143] Another feature is the automatic calculation of quality metrics directly from the guidances and protocol deviation system embedded into the system. This data is constantly calculated and collected by the system and can be easily tracked by shift, day / night, clinician, unit, or hospital - providing a rich set of quality metric data virtually for free.

[0144] Regulatory Integration Framework

[0145] The platform seamlessly integrates with regulatory requirements through: (1) Joint Commission Reporting: The system aligns with The Joint Commission’s “gold standard” in performance measurement, automatically formatting ventilator or device care metrics according to their Specifications Manual for National Hospital Inpatient Quality Measures to minimize duplicate data collection efforts. (2) FDA Quality Surveillance: The platform would support FDA’s Quality Metrics Reporting Program by providing quantitative and objective measures of ventilator management quality that can be integrated into comprehensive quality surveillance programs. (3) HIPAA / GDPR Compliance: The system incorporates built-in protections including: Data privacy controls with encryption and access restrictions, Technical safeguards for electronic health record protection, Explicit consent management for data processing, and Purpose limitation to ensure data is used only for authorized purposes.

[0146] By transforming the device alarm management platform into a comprehensive quality metrics system, healthcare providers could move from reactive compliance documentation to proactive quality improvement while significantly reducing the administrative burden of regulatory reporting.

[0147] Non-Limiting Examples

[0148] All references cited herein are expressly incorporated by reference in their entirety. It will be appreciated by persons skilled in the art that the present disclosure is not limited to what has been particularly shown and described herein above. In addition, unless mention was made above to the contrary, it should be noted that all of the accompanying drawings are not to scale. There are many different features to the present disclosure and it isAttorney Docket No. 9224-P0007-WO 28contemplated that these features may be used together or separately. Thus, the disclosure should not be limited to any particular combination of features or to a particular application of the disclosure. Further, it should be understood that variations and modifications within the spirit and scope of the disclosure might occur to those skilled in the art to which the disclosure pertains. Accordingly, all expedient modifications readily attainable by one versed in the art from the disclosure set forth herein that are within the scope and spirit of the present disclosure are to be included as further embodiments of the present disclosure.

[0149] The description in the present application should not be read as implying that any particular element, step, or function is an essential or critical element that must be included in the claim scope. The scope of patented subject matter is defined only by the allowed claims. Moreover, none of the claims invokes 35 U.S.C. § 112(f) with respect to any of the appended claims or claim elements unless the exact words “means for” or “step for” are explicitly used in the particular claim, followed by a participle phrase identifying a function.

[0150] Benefits, other advantages, and solutions to problems have been described above with regard to specific embodiments. However, the benefits, advantages, solutions to problems, and any feature(s) that can cause any benefit, advantage, or solution to occur or become more pronounced are not to be constructed as a critical, required, sacrosanct, or an essential feature of any or all the claims.

[0151] After reading the disclosure, skilled artisans will appreciate that certain features are, for clarity, described herein in the context of separate embodiments, can also be provided in combination in a single embodiment. Conversely, various features that are, for brevity, described in the context of a single embodiment, can also be provided separately or in any sub -combination. Further, references to values stated in ranges include each and every value within that range.

[0152] The above discussion is meant to be illustrative of the principles and various embodiments of the present invention. Numerous variations and modifications will become apparent to those skilled in the art once the above disclosure is fully appreciated. It is intended that the following claims be interpreted to embrace all such variations and modifications.

[0153] Incorporated References[1] Russo, E., et al., Challenges inpatient safety improvement research in the era of electronic health records. Healthc. (Amst), 2016. 4(4): p. 285-290.[2] Balas, E.A., et al., Big Data Clinical Research: Validity, Ethics, and Regulation. Stud. Health Technol. Inform, 2015. 216: p. 448-452.[3] Raghupathi, W. and V. Raghupathi, Big data analytics in healthcare: promise and potential. Health Inf. Sci. Syst, 2014. 2: p. 3.[4] Sittig, D.F. and H. Singh, Electronic health records and national patient-safety goals. N. Engl. J. Med, 2012. 367(19): p. 1854-1860.[5] Olchanski, N., et al., Can a Novel ICU Data Display Positively Affect Patient Outcomes and Save Lives? J Med Syst, 2017. 41(11): p. 171.[6] Waller, R.G., et al., Novel displays of patient information in critical care settings: a systematic review. J Am Med Inform Assoc, 2019. 26(5): p. 479-489.Attorney Docket No. 9224-P0007-WO 29[7] Lee, M.; Aswani,A., Methodology for Interpretable Reinforcement Learning for OptimizingMechanical Ventilation. 2024.

[0154] What is claimed is:Attorney Docket No. 9224-P0007-WO 30

Claims

CLAIMS1. A method for managing respiratory ventilators, the method comprising:communicatively coupling one or more respiratory ventilators to at least one computing system, each of the one or more respiratory ventilators provides ventilator-related data associated with a ventilated patient, wherein the ventilator-related data includes one of ventilator settings, ventilator alarms, ventilator measured parameters or a combination thereof;receiving, by the at least one computing system, patient-related data in a first data format including data from the one or more respiratory ventilators, wherein the patient-related data includes one of the ventilator-related data, physiologic data, demographic data, diagnosis data, laboratory data, or a combination thereof, which is associated with the ventilated patient;converting, by the at least one computing system, the patient-related data in the first data format into a second standardized data format;providing, over a network, access to a graphical user interface (GUI) configured to display at least a portion of the patient-related data in substantially real time;receiving updated patient-related data from at least one of the one or more respiratory ventilators; converting the updated patient-related data into the second standardized data format to generate standardized updated patient-related data;automatically updating the GUI in response to receipt of the standardized updated patient-related data; creating a system prompt in substantially real time for an artificial intelligence (Al) system using the updated patient-related data;in response to creating the system prompt, generating by the Al system one or more recommendations for improving care of the ventilated patient; andapplying at least one of the one or more recommendations to the at least one of the one or more respiratory ventilators associated with the updated patient-related data.

2. The method of claim 1, wherein the Al system is instructed by the system prompt or by design to use a curated library of source materials, including one of clinical practice guidelines (CPG), hospital protocols, clinical literature, or a combination thereof; andin response to the one or more recommendations being within a given settable threshold, automatically updating the GUI with the updated ventilator setting recommendations.

3. The method of claim 2, wherein the source materials used by the Al system further comprises using a combination of the curated library of source materials with one of knowledge graphs, decision trees, flow charts, or a combination thereof, to generate more deterministic and verifiable ventilator setting recommendations.

4. The method of claim 3, wherein the knowledge graphs, decision trees, or flow charts are extracted automatically from clinical practice guidelines, hospital protocols, or medical literature by LLMs as part of the Al system.Attorney Docket No. 9224-P0007-WO 315. A respiratory ventilator management system comprising:one or more respiratory ventilators configured to generate ventilator-related data associated with a ventilated patient, wherein the ventilator-related data includes one or more of ventilator settings, ventilator alarms, ventilator parameters, or a combination thereof, which is associated with the ventilated patient;at least one computing system communicatively coupled to the one or more respiratory ventilators, the at least one computing system comprising one or more processors and memory storing instructions that, when executed by the one or more processors, cause the at least one computing system to:receive patient-related data in a first data format including data from the one or more respiratory ventilators, wherein the patient-related data includes one of the ventilator-related data, physiologic data, demographic data, diagnosis data, laboratory data, or a combination thereof, which is associated with the ventilated patient;convert the patient-related data in the first data format into a second standardized data format; provide, over a network, access to a graphical user interface (GUI) configured to display at least a portion of the patient-related data in substantially real time;receive updated patient-related data from at least one of the one or more respiratory ventilators; convert the updated patient-related data into the second standardized data format to generate standardized updated patient-related data;automatically update the GUI in response to receipt of the standardized updated patient-related data;create a system prompt in substantially real time for an artificial intelligence (Al) system using the updated patient-related data;provide the system prompt to the Al system;receive, from the Al system and in response to the system prompt, one or more recommendations for improving care of the ventilated patient; andapply at least one of the one or more recommendations either directly or through a clinician to at least one of the one or more respiratory ventilators associated with the updated patient -related data.

6. The respiratory ventilator management system of claim 5, wherein the Al system is instructed by the system prompt or by design to use a curated library of source materials, including one of clinical practice guidelines (CPG), hospital protocols, clinical literature, or a combination thereof; andin response to the one or more recommendations being within a given settable threshold, automatically update the GUI with the updated ventilator setting recommendations.

7. The respiratory ventilator management system of claim 6, wherein the source materials used by the Al system further comprise using a hybrid model that combines the curated library of source materials with one of knowledge graphs, decision trees, flow charts, or a combination thereof, to generate more deterministic and verifiable ventilator setting recommendations.

8. The respiratory ventilator management system of claim 7, wherein the knowledge graphs, decision trees, or flow charts are extracted automatically from clinical practice guidelines, hospital protocols, or medical literature by LLMs as part of the Al system.Attorney Docket No. 9224-P0007-WO 329. The respiratory ventilator management system of claim 5, further comprising:creating the system prompt or designing the Al system to use a curated library of source materials, the source materials including a natural language summary of clinical practice guidelines (CPG), hospital protocols, clinical literature, or a combination thereof;wherein the natural language summary is configured to improve interpretability of the source materials by the Al system to create more accurate and reliable recommendations from the Al system; andwherein the natural language summaries are configmed for clinician verification.

10. The respiratory ventilator management system of claim 9, wherein the natural language summary contains specific rules for specific patient conditions, and the system prompt includes instructions on how to determine the specific patient conditions.

11. The respiratory ventilator management system of claim 9, wherein the system prompt further includes instructions causing the Al system to estimate a patient condition using one of the standardized updated patient-related data, patient demographic data, diagnosis data, laboratory data, or a combination thereof to estimate the patient condition.

12. The respiratory ventilator management system of claim 6, wherein the system prompt includes patientspecific conditions including one of COPD, ARDS, is obese, or a combination thereof, and using these patientspecific conditions along with the curated library of source materials to determine patient-specific ventilator setting recommendations.

13. The respiratory ventilator management system of claim 5, further including one or more of the system prompt features to improve the one or more recommendations for clinical decision support:instructions to format the one or more recommendations in an easily readable format for clinicians, instructions to format a tone of the one or more recommendations such that concise recommendations are described first for quick review by the clinician, followed by detailed rationale and references for the concise recommendations,instructions to calculate specific parameters that enhance patient-specific recommendations, hypertext or similar markup language formatting allowing entry of data into the system prompt at specified locations,instructions to ignore or replace data that is older than a settable parameter, ora combination thereof.Attorney Docket No. 9224-P0007-WO 3314. A respiratory ventilator management system comprising:one or more respiratory ventilators configured to generate ventilator-related data associated with a ventilated patient, wherein the ventilator-related data includes one or more of ventilator settings, ventilator alarms, ventilator parameters, or a combination thereof, which is associated with the ventilated patient;at least one computing system communicatively coupled to the one or more respiratory ventilators, the at least one computing system comprising one or more processors and memory storing instructions that, when executed by the one or more processors, cause the at least one computing system to:receive patient-related data in a first data format from the one or more respiratory ventilators; convert the patient-related data in the first data format into a second standardized data format; provide, over a network, access to a graphical user interface (GUI) configured to display at least a portion of the patient-related data in substantially real time;receive updated patient-related data in the first data format from at least one of the one or more respiratory ventilators;convert the updated patient-related data in the first data format into the second standardized data format to generate standardized updated patient-related data;automatically update the GUI in response to receipt of the standardized updated patient-related data;comparing the updated patient-related data to one of clinical practice guidelines (CPG), hospital protocols, or both;designate, on the GUI, any of the updated patient-related data that does not comply with the one CPG, hospital guidelines, clinical literature, or a combination thereof;provide, on the GUI, one or more recommendations on how to bring the updated patient -related data back into compliance with the one CPG, hospital guidelines, clinical literature, or a combination thereof; andapply at least one of the one or more recommendations either directly or through a clinician to at least one of the one or more respiratory ventilators associated with the updated patient -related data.

15. The respiratory ventilator management system of claim 14, wherein the comparing the updated patient-related data with one of clinical practice guidelines (CPG), hospital protocols, or both, further comprising:automatically converting the updated patient-related data and user input into a customizable shift report format for one or more patients; anddisplaying, on the GUI, the customizable shift report to the user during shift change communication.

16. The respiratory ventilator management system of claim 15, further including:receiving from a clinician, via the GUI, additions and revisions to the customizable shift report format; andwherein these additions and revisions include one or more of motivation or rationale for ventilator changes, clinical recommendations for the next shift, patient response to treatments, procedures completed during the shift, or procedures or events expected in the next shift, or a combination thereof.Attorney Docket No. 9224-P0007-WO 3417. The respiratory ventilator management system of claim 14, further comprising:storing the updated patient-related data in the second standardized data format in a data repository associated with the at least one computing system, wherein the storing includes storing with each data item at least one timestamp that indicates when the updated patient-related data was retrieved from one of the respiratory ventilators;calculating a time-since parameter for each data item, wherein the time-since parameter indicates a duration of time that has elapsed between the time recorded in the at least one timestamp and a current time; and providing access over a network to the at least one computing system for users to view at least some of the updated patient-related data from the one or more respiratory ventilators in substantially real time through a graphical user interface (GUI).

18. The respiratory ventilator management system of claim 17, wherein the time-since parameter includes one of:using an indicator of the updated patient-related data displayed on the GUI based on time; disabling of recommendations of ventilator settings that are dependent on updated patient -related data that is older than a settable time threshold;enabling a further level of recommendations that are triggered when conditions have existed for greater than a settable time threshold;calculating how often key measurements are taken and using this information for quality improvement, the key measurements include one of blood gases, baseline measurements, procedures, or a combination thereof; or a combination thereof.

19. The respiratory ventilator management system of claim 17, wherein the updated patient-related data, the time-since parameter, and the at least one timestamp are used to temporarily align updated patient -related data, wherein a portion of the updated patient-related data has a long delay between measurement and results being available compared with older real-time data.

20. The respiratory ventilator management system of claim 17, further comprising:using multi-scale averaging of updated patient-related data to identify trends and values that change over time, to enable more complex recommendations.Attorney Docket No. 9224-P0007-WO 35