DETERIORATION CASE STUDY – DRAFT
[SECTION] – THE CHALLENGE:
Identifying and Impacting All- Cause Deterioration
The healthcare industry faces a significant challenge in identifying and impacting patient deterioration, which is a common occurrence during hospital stays. Thousands of patients experience unexpected clinical deterioration every year, resulting in unplanned ICU transfers or even death. To improve patient outcomes, early detection of deterioration is crucial, and this requires timely and accurate models. However, given the heterogeneity of deterioration and the different reasons for ICU transfers, the healthcare industry needs both general and condition-specific models to effectively identify and impact all-cause deterioration.
Addressing these challenges can lead to improved patient outcomes and a more efficient healthcare system.
[SECTION] – HIGH LEVEL DESCRIPTION
Overview of Bayesian’s clinical AI solution: Provide a brief introduction to the solution, including what it is designed to do, how it works, and what makes it unique.
[SECTION] – USE CASE
Describe the specific all-cause deterioration use case that the solution addresses, including the key challenges or pain points that healthcare providers face in managing this condition.
[SECTION] – TARGET AUDIENCE
Identify the healthcare providers who would benefit most from the solution, such as clinicians, nurses, or care managers.
[SECTION] – CLINICAL BENEFIT
Explain how the clinical AI solution helps healthcare providers address the challenges of the all-cause deterioration use case, including improving patient outcomes, reducing costs, or increasing efficiency.
[SECTION] – OUR APPROACH
Bayesian’s Approach to Clinical AI
Bayesian’s Clinical AI platform uses both structured and unstructured data to ingest a vast variety of data in the patient’s record, including labs, vitals, patient medical history, treatments, medications, clinical notes, location, history, problem list, and more.
In addition, Bayesian uses high-quality, clinical-grade phenotypes to reason about a patient’s risk status, allowing for more precise and accurate insights. Our learners use multiple strategies, including problem-specific strategies, to tackle data complexity and unique clinical areas. This allows for greater accuracy and better insights for clinicians, ultimately resulting in improved patient outcomes.
Bayesian is tuned to specific patient populations because patients in different areas of the country can have vastly different healthcare needs. Furthermore, Bayesian is also tuned to specific clinical circumstances based on a patient’s physiology, resulting in more personalized insights and improved accuracy.
As a result of this approach, Bayesian’s resulting models are accurate, precise, and actionable, reducing false alerting, improving precision, improving lead time, and improving actionability. Bayesian delivers insights that clinicians can trust, resulting in improved patient outcomes and better overall healthcare delivery.
Focus on Adoption….. Behavior change, trust, clinical colleague/can disagree
Augmentation… Capacity gain
AI DONE RIGHT [SECTION]
Improved Positive Predictive Value
With a 90% sensitivity rate, our process can improve PPV by a staggering 70%, while a sensitivity rate of 80% can still achieve an impressive improvement of 40%.
Utilizing Richer Inputs and High-Quality Targets to Accurately Predict Patient Outcomes
By combining richer inputs with high-quality targets, we can accurately predict patient outcomes with a 106% increase in area under the precision-recall curve, showcasing our process as the most effective approach for detecting patient deterioration.
Advanced Condition-Specific Models
Bayesian Health has revolutionized the early detection and prevention of diseases with our advanced condition-specific models. Our models for deterioration were developed by learning combinations of clinical markers and their trends that are predictive of patients who transfer to the ICU from the floor or the step-down unit or die within their hospital stay. By capturing early signs of common causes of unplanned ICU transfers, such as sepsis, acute respiratory failure, GI bleed, shock, and cardiac arrest, Bayesian helps clinicians intervene early and improve patient outcomes
Early Detection Leading to Better Outcomes
By providing a median warning of up to 30 hours before onset, our models can detect potential organ dysfunction and give clinicians ample time to intervene. Furthermore, our models interact with our deterioration alert system, allowing clinicians to address sepsis early and prevent ICU transfer, or accelerate sepsis bundle when deterioration alerts come before sepsis alerts.
Smart Interpolation Techniques
With our advanced smart interpolation technique and Wait and Watch policy, we accurately predict outcomes despite missing data, reducing false alerts by up to 3-4x. Trust Bayesian Health to deliver unmatched insights and improve patient outcomes through our innovative and advanced predictive models.
To monitor patient care, three different views are used: system-level adherence, case-level abstraction, and phenotyping. The system-level adherence view tracks every step in an interaction, the case-level abstraction view reviews information from the prior day, and the phenotyping view identifies patients who are deteriorating and in need of critical care. These views provide insight into engagement, specific workflow steps, and areas for improvement.
High-Quality Targets and Rich Inputs
To improve patient outcomes, high-quality targets, such as labs, vitals, notes, orders, medications, chief complaints, history, and deviation from baseline, are used to identify primary signs of infection and infection-associated acute organ dysfunction. Confounding comorbidities are explained by other factors, and richer inputs and high-quality targets improve positive predictive value (PPV) and the area under the precision-recall curve.
Models to detect sepsis and acute respiratory failure are tuned to patient sub-populations and care units. The sepsis model has a 94% sensitivity and 30-40% PPV, while the acute respiratory failure model has a 90% sensitivity and 40% PPV. The interaction between sepsis and deterioration models provides early attention to prevent ICU transfer or accelerate the sepsis bundle.
Accounting for Bias
Bayesian’s models are accurately tuned to specific patient populations and have a robust monitoring system for bias and dataset shift, ensuring accuracy and reliability, regardless of the patient population. The platform monitors for bias in terms of assessing model performance as well as looking for indicators of bias in clinical practice patterns. By using different strategies to account for bias, such as enriching data sets by finding additional data to clean it up, Bayesian enables hospitals to learn patterns that are less biased towards the practices of a single health system and more general about disease, health, and medicine.
End-to-End Performance Optimization
Bayesian Health’s platform uses machine learning to enable clinicians to act on patient-specific insights in their workflow. The program is applicable across acute, post-acute, and virtual care problems and aims to ensure early detection and warning, best practice, and treatment adherence. It is designed to create sustained engagement and adoption and is built on behavior change principles. The program includes a clinical discovery process, site-specific rollout, multi-level monitoring, and analysis to identify gaps and optimization support.
[SECTION] – ALERTING DONE RIGHT
Alerting Within the EMR
Bayesian’s ability to deliver actionable, contextualized clinical alerts within the EMR is crucial in ensuring adoption of the technology. By integrating within a clinician’s workflow in the EMR and partnering with leading EMR systems, Bayesian provides comprehensive condition alerting, monitoring, and treatment tracking, allowing users to stay within the module for as much of the condition care pattern as possible. This minimizes clicks and additional effort, maximizing efficiency and clinical efficacy. Additionally, Bayesian’s novel machine learning infrastructure provides high-quality, precise, and patient-specific inferences at the right time, resulting in far fewer false alerts and reducing alarm fatigue.
Alerting: Clinical Risk Stratification
Bayesian provides two types of deterioration alerts: moderate risk and high risk, enabling clinicians to take action based on the patient’s level of deterioration. Moderate risk alerts are highly sensitive and notify clinicians early, while high risk alerts are highly specific and caution providers when a patient might require critical attention. Patients flagged as moderate risk have an 8x higher risk for ICU transfer and 7x higher risk of mortality. Patients flagged as high risk have a 56x higher risk for ICU transfer than patients not flagged as high risk.
[SECTION] – IMPLEMENTATION DONE RIGHT
End-To-End Clinical & Operational Deployment
Bayesian Health offers an end-to-end Client Success program that simplifies the deployment of their AI models and technology, ensuring a streamlined process that meets the unique needs and goals of each organization. By working closely with stakeholders within the organization, Bayesian Health optimizes existing processes to create a comprehensive implementation plan, and then monitors progress throughout the deployment. The Client Success team also performs ongoing monitoring, tuning, and performance improvement measures post-deployment to ensure the long-term efficacy and effectiveness of the investment in Bayesian Health. Bayesian’s implementation approach includes best practice workflows and specific configurability points, resulting in a customizable and streamlined deployment process. With the use of cutting-edge technology and expertise, Bayesian Health strives to achieve rapid success in all-cause deterioration, lower mortality rates, and augment caregivers with high-quality clinical AI.
Bayesian designs and deploys a user-friendly, intuitive, and efficient workflow integration and access process by understanding the provider and nurse workflow for the care units where the solution will be deployed. By following these steps, Bayesian ensures that its solution meets the needs of healthcare providers and patients.
To determine the necessary workflow configurations for deploying a module, Bayesian maps out the current and future state by gathering details on the client’s clinical user workflow, presenting a workflow recommendation to key clinical groups, and defining a test plan for each module. Client Success collaborates with internal stakeholders in the creation of a technical integration roadmap and builds comprehensive and customizable reports to monitor adoption rates and system performance, identify areas requiring improvement, and take corrective action.
[SECTION] – RESEARCH/EVIDENCE OF EFFICACY
Provide data or other evidence that demonstrates the efficacy of the solution, such as clinical trial results or real-world case studies.
[SECTION] – ROI
Return On Investment
By implementing Bayesian’s clinical AI system, hospitals can benefit from increased efficiency and a significant return on investment.
For an average 200 bed hospital, we’ve calculated the financial benefit and capacity gain based on rigorous assumptions from literature, including Bayesian’s performance in published literature. The results are astounding:
- Financial Benefit: ~$2-3M financial benefit Reduced utilization and length of stay Revenue gain due to new bed capacity Reduction in HAC penalties Avoidance of malpractice lawsuits Automation of POA documentation to drive revenue capture
- Capacity Gain: 3-6 FTE capacity gain Time saved from fewer alerting Time saved from automating case reviews through reporting Time saved from automated documentation of accurate coding language
With Bayesian, hospitals can reduce utilization and length of stay, which results in cost savings and increased revenue. Additionally, by automating documentation, hospitals can ensure that they are accurately capturing revenue, avoiding penalties, and reducing the risk of malpractice lawsuits.
Moreover, hospitals can gain additional capacity through increased efficiency, with fewer alerts and automated reporting, enabling clinicians to spend more time providing care to patients. This results in a significant return on investment, with a financial benefit of up to $2-3M and a capacity gain of 3-6 FTEs.
[SECTION] – CONCLUSION
Unmatched Insights and Innovative Approach to Improving Patient Outcomes
Accurately predicting patient outcomes is critical for better patient care. General and condition-specific models, along with multi-level visibility and Bayesian Health’s End-to-End Performance Optimization Program, can improve patient outcomes. Bayesian Health’s program provides a flexible framework for ingesting new sources and creating accurate, precise, and specific models. The program’s intuitive user interface and adoption strategy create sustained engagement and adoption. By achieving a balance between sensitivity and specificity in identifying and impacting deterioration, healthcare providers can provide better patient care.
Identifying and impacting deterioration in patients is a complex task that requires timely and accurate models. The heterogeneity of deterioration, along with the various reasons for unplanned ICU transfers, necessitates the use of general and condition-specific models. Early detection is critical in achieving better patient outcomes, but excessive alerts can be ignored. Studies have shown that the National Early Warning Score (NEWS) has poor performance characteristics and is generally ignored by frontline nursing staff.
To address these challenges, two alerts are proposed: the yellow alert, which is highly sensitive and fires early, and the red alert, which is highly precise and specific and alerts physicians immediately. Both alerts rely on the same underlying model, with the yellow alert providing a 7x increased risk of ICU admission and mortality and a median early warning of 24 hours, while the red alert provides a 7x increased risk of mortality and a median early warning of 48 hours. A training dataset of 220k encounters was used to identify positive outcomes and tune the alerts for positive predictive value and sensitivity. In addition, condition-specific models supplement the deterioration model to provide more accurate predictions. Multi-level visibility drives performance optimization by providing insight at each stage of care progression, from identifying high-risk patients to treatment completion. Metrics include interaction summaries, response timing, system-level adherence, and provider engagement, with visibility into adherence at the individual provider level. The goal is to achieve a balance between sensitivity and specificity in identifying and impacting deterioration, resulting in better patient outcomes.
three different views for monitoring patient care: system-level adherence, case-level abstraction, and phenotyping. The system-level adherence view tracks every step in an interaction in each care unit to identify gaps in engagement and view specific workflow steps. The case-level abstraction view allows care unit leaders to review information from the prior day to ensure that at-risk patients were responded to appropriately and that necessary interventions were given in a timely manner. The phenotyping view aims to identify patients who are deteriorating and in need of critical care by using a proxy measure such as unplanned ICU transfers or in-hospital deaths. Data used includes labs, vitals, age, and gender, with a look back of up to 6 hours for vitals and up to 2 days for labs. Although these views have their pros and cons, they are important tools for healthcare providers to monitor patient care and identify areas for improvement.
Bayesian Health’s End-to-End Performance Optimization Program is a platform that uses best-in-class machine learning to enable clinicians to act on patient-specific insights in their workflow. The platform is built on behavior change principles and is designed to create sustained engagement and adoption. The program is applicable across acute, post-acute, and virtual care problems and aims to ensure early detection and warning, best practice, and treatment adherence. The platform is integrated into the EMR to provide comprehensive condition alerting, monitoring, and treatment tracking. The platform’s flexible framework for ingesting new sources and best-in-class target identification algorithms and learners create accurate, precise, and specific models. The program includes a clinical discovery process, site-specific rollout, multi-level monitoring, and analysis to identify gaps and optimization support. The platform’s intuitive UI, positive feedback loop, and engagement with stakeholders and users ensure sustained adoption and a habit of use. The program is designed to activate buy-in and create a habit of use, measure and tune areas for optimal performance, and align with stakeholders on goals and success metrics.
Confounding comorbidities such as high creatinine and high WBC can be explained by other factors like CKD and recent surgery. Labs, vitals, notes, orders, medications, chief complaints, history, and deviation from a patient’s baseline can be used to identify primary signs of infection and infection-associated acute organ dysfunction. High quality targets have a 90% sensitivity and can improve PPV by 70% while high quality targets with 80% sensitivity can improve PPV by 40%. In contrast, poor targets are not effective in detecting patient deterioration.
Combining richer inputs with high-quality targets can significantly improve patient outcome predictions. The data shows that at 90% sensitivity, there is a remarkable 310% improvement in positive predictive value (PPV), and at 80% sensitivity, there is an impressive 180% improvement in PPV. Rich inputs and high-quality targets result in a 106% increase in area under the precision-recall curve, a measure of the model’s performance. At 80% recall, PPV increases by 178% and at 90% recall, PPV increases by 310%, while at 95% recall, PPV increases by an impressive 500%. This suggests that using richer inputs and high-quality targets is the most effective approach in accurately predicting patient outcomes. In contrast, structured inputs with poor targets are not effective in detecting patient deterioration.
Condition-specific models are essential to supplement the deterioration model, and they can significantly improve early detection and prevention of diseases. For instance, models have been created to detect sepsis and acute respiratory failure, and they are tuned to patient sub-populations and care units. The sepsis model has a 94% sensitivity and 30-40% positive predictive value (PPV) and can provide an early warning before organ dysfunction occurs. The acute respiratory failure model, on the other hand, has a 90% sensitivity and 40% PPV and can give a median warning of 24-30 hours before the onset of ARF. Interestingly, there is also an interaction between the sepsis and deterioration models. When the sepsis alert comes before the deterioration alert, which happens about 50% of the time, addressing sepsis early could prevent ICU transfer. Conversely, when the sepsis alert comes after the deterioration alert, early attention could accelerate the sepsis bundle. These models’ specificity and sensitivity provide valuable insights into patient care, which can lead to improved outcomes.
Clinical data is often missing or collected at different times and frequencies for patients, making it difficult to accurately predict outcomes. To address this challenge, a smart interpolation technique based on patient context is used to account for missing data. The Wait and Watch policy is also implemented, which balances the delay in making predictions against the cost of false alerts. The method involves predicting septic shock based on 10 longitudinal clinical inputs. By using a smart learner that accounts for missing data, false alerting is significantly reduced, resulting in a 3-4x improvement in the number of patients detected at the same false alerting rate compared to using a standard approach like logistic regression. This demonstrates the importance of accounting for missing data in predictive models and how a smart learner can improve the accuracy of patient outcome predictions.
The engine, when combined with Optimization Support, can identify gaps and score opportunities for improvements in various areas such as staff use, best practice adherence, and documentation. This leads to actionable improvement opportunities for targeted education, optimizing bundles and pathways, and improving billing and coding. After launch, the performance monitoring and optimization feature facilitates sustained adoption through adherence monitoring, automated case abstraction, and outcomes monitoring. Multi-level monitoring is necessary to identify and monitor engagement and improve adoption by targeting gaps and creating targeted action and education. Comprehensive insights and radical transparency are crucial for trust and sustained use by identifying gaps and drilling down from the aggregate to individual cases. The insights generated can then be used to identify actionable opportunities for care pathway improvement, targeted education, and streamlining billing and coding processes.
the challenges in detecting patient deterioration in hospitals and the need for a more precise and actionable approach. ICU transfers are expensive and difficult to predict, and deterioration is highly related to patient mortality. Patient heterogeneity, complex illness presentations, and different needs for different user types make it challenging to identify deterioration in a timely and effective manner. The article proposes a solution that incorporates multiple highly-tuned models and real-time insights for performance optimization, adoption, and quality initiatives. The approach focuses on precision predictive analytics and a clinically informed approach to early detection of illnesses before they become critical.
several challenges in providing enhanced care in hospitals and proposes various solutions to address these challenges. The first challenge is early identification, which is crucial for accurate documentation and preventing the progression of illnesses. The article highlights a thorough and high-quality early detection model that has shown promising results.
The second challenge is the inconsistency in the application of workflows, which leads to scattered data. The article proposes easy and interpretable evaluations, longitudinal monitoring, and real-time views to ensure consistency in the application of interventions.
The third challenge is the variability in nursing interpretation, which makes it difficult to tailor interventions. The article proposes using natural risk scoring and braden to grade the degree of escalation, which may require FIGMA modification.
The fourth challenge is scaling up QI, which spends an enormous amount of time doing chart abstraction. The article proposes a central QI team that can easily monitor patients at a glance, granular monitoring to provide reports for patients missed and consistency of application, and gap identification to facilitate where to focus on problematic units and potential for optimizing bundle recommendations by risk.
The fifth challenge is improved coding and billing, which naturally flows out of improved documentation. The article proposes additional reports available for the coding team to do period checks to ensure that all cases with POA Y have appropriate documentation.
The article discusses statistical models for early detection of subacute potentially catastrophic illnesses that lead to ICU transfer from a cardiac and cardiac surgery ward. The study found that phenotypes vary among clinical conditions that lead to ICU transfer and that statistical models trained specifically for each condition perform better than a one-size-fits-all approach. The addition of continuous cardiorespiratory monitoring is recommended to fill in the timeline between nurse visits and blood draws. The article proposes “precision predictive analytics monitoring” as a more focused and clinically informed approach to early detection of subacute potentially catastrophic illnesses in hospitalized patients. The study’s strength lies in the individual review of charts to identify reasons for ICU transfer. Finally, the need to look for multiple modes of critical illness is emphasized, as no single predictive model seems adequate for sepsis.