AI/ML Models & ExplainerAI Dashboards

Pediatric Acute Chest Syndrome Predictor (ACS-PS-RISK)

Problem

Acute Chest Syndrome (ACS) presents a significant health challenge in pediatric sickle cell disease (SCD), accounting for 15-30% of all pediatric SCD hospital admissions. Each episode of ACS not only increases the risk of morbidity and mortality but also poses a threat of irreversible lung damage. Despite known risk factors such as a history of asthma, ACS often arises unexpectedly, complicating patient management and increasing the burden on healthcare systems. The unpredictability of ACS exacerbates the need for enhanced monitoring and preventive strategies in hospital settings, leading to increased use of medical resources and extended hospital stays. 

Solution

This advanced Machine Learning model excels in predicting ACS with high accuracy in a pediatric cohort. By identifying high-risk patients early, we can significantly enhance treatment options and improve inpatient management strategies, ensuring better patient outcomes.

Outcome

Our state-of-the-art random forest model predicts ACS 24 hours before chart diagnosis with an impressive sensitivity of 82.5% and a specificity of 61.3% at a 0.17 cutoff point. It boasts a remarkable 93% negative predictive value and a 34% positive predictive value.

Service Line

Pediatrics

Risk Score to Predict Readiness for Surgery (ASA-PS-RISK)

Problem

The American Society of Anesthesiologists Physical Status (ASA-PS) classification system, established in 1941, is instrumental in evaluating a patient’s pre-anesthesia medical comorbidities. However, its application has revealed significant challenges, primarily due to its reliance on subjective human judgment. Variability in assessments among anesthesiologists and consistent underestimation by surgeons and their teams can lead to inaccurate patient classifications. This inconsistency can adversely affect the allocation of healthcare resources and the planning of preoperative evaluations, potentially leading to unnecessary testing, operative delays, or even cancellations on the day of surgery. As such, there is a clear need for a more reliable method to assign ASA-PS classifications to enhance the accuracy and effectiveness of perioperative risk management.

Solution

We developed a machine learning physical status model using preoperative data, improving early identification of high-risk patients. This model standardizes preoperative evaluation, enhancing outcomes for ambulatory surgery patients.

Outcome

Our study showed that the anesthesiologist ASA-PS and ML-PS agreed 57.2% of the time. The ML-PS more often assigned patients to extreme ASA-PS categories (I and IV) and fewer to II and III (p<0.01). Both systems had excellent predictive values for 30-day mortality and good values for postoperative ICU admission and adverse discharge. Among 3,594 patients who died within 30 days post-surgery, ML-PS reclassified 1,281 (35.6%) into a higher risk category than anesthesiologists did. However, for patients with multiple co-morbidities, anesthesiologist-ASA-PS had better predictive accuracy.

Service Line

Perioperative

Published Outcomes

ScienceDirect

Acute Respiratory Distress Syndrome Predictor (ARDS-RISK)

Problem

Acute respiratory distress syndrome is dangerous, especially in critical care and was the primary cause of death for COVID patients

Solution

The early recognition ML model predicts ARDS risk factors and identifies patients, allowing care teams to intervene sooner and prevent poor or worsening outcomes.

Service Line

Respiratory

Published Outcomes

NCBI

Automatic Chart Abstraction and CPT Code Generation (AUTOCHART-LLM)

Problem

Hospitals are mandated by regulatory and standards bodies to report quality and outcomes data. Signifiant time and expense is spent annually by hospitals on nursing abstraction services for purposes of reimbursement, board certifications, value-based care and population health reporting.

However, manual chart review by health-system quality teams and nurses is both costly and inefficient for several reasons. First, it requires significant labor, with skilled professionals spending extensive time examining and interpreting large volumes of patient records. Second, the process is prone to human error, which can lead to inconsistencies in data extraction and analysis. Third, it lacks scalability, as increasing patient volumes directly translate to increased labor and costs. Finally, manual reviews often result in delayed insights, slowing down the implementation of necessary quality improvements and adjustments in patient care practices.

Solution

Our innovative solution enhances manual chart review by utilizing a fine-tuned Large Language Model to automatically extract and link key concepts, like CPT codes, from complex medical notes, streamlining the process and ensuring 94%+ accuracy. The model can be tuned to search for any key concept. See our Clinical Trials Matching LLM for another example of this technology in use.

Outcome

Nurse data abstraction from patient charts, determining the appropriate CPT codes from clinical notes and data entry into  portals or registries for numerous quality and performance reporting needs creates significant operational overhead for health systems. Using our AUTOCHART LLM, sometimes as part of an ensemble with our other LLMs, can fully automate most chart abstraction needs.

In one use case, the ACS National Surgical Quality Improvement Program, our LLM and ETL solution 100% eliminated data abstraction/extraction and data entry time, auto-generated CPT codes from clinical notes and automatically entered discrete data into the appropriate fields within the ACS registry, leaving only the last step of validation to be performed by staff. 5 year projections show a potential for $700K-$1M in cost savings, significant ROI and better utilization of depleted nursing resources.

Service Line

Quality, Revenue Cycle, Population Health, Accreditation

De-Identification & PHI Redaction LLM (DEID-LLM)

Problem

In daily health-system operations, there are various groups that need to see clinical text generated by treatment teams, but to ensure privacy and regulatory compliance may only see this data in a de-identified manner. Examples include research, billing, revenue cycle, QI/PI, population health and various other operational or reimbursement purposes and, especially, where an outside entity is the one reviewing the data.

Solution

The DEID-LLM allows for real-time, large-scale de-identification of clinical text. Once processed, the de-identified text can be immediately utilized by downstream workflows, internally or externally of institution, as required. Integrated within a business process workflow, this LLM has the potential to produce significant ROI for health systems. A potential solution for cybersecurity, privacy and compliance teams or those teams which have data governance or oversight responsibilities.

Service Line

Hospital-Wide

Case Cancelation: Outpatient Visit No-Show Prediction

Problem

For health systems, missed appointments in ambulatory settings represent a significant challenge. Every time a patient fails to show up for a scheduled visit, a valuable appointment slot goes unused. This not only leads to lost revenue for healthcare providers but also denies urgent care to other patients who could have benefited from that slot. Moreover, patients who miss their appointments miss out on critical care opportunities, compromising their health outcomes.

Solution

Building on our Surgical Case Cancellation model, Cognome developed a Machine Learning algorithm to predict outpatient visits no-shows and patient cancellations, in order to allow health-system to optimize available appointment time-slots.

Service Line

Outpatient, Ambulatory

Case Cancellation: Surgical Same Day Cancellation

Problem

Solution

Cognome developed a Machine Learning model to predict case cancellations within 24 hours of surgery. This model utilizes insights from 29 predictors, integrated into the EHR, allowing case managers to proactively address medical, behavioral health, transportation or SDOH issues in order to optimize clinician and resource utilization.

Outcome

This model is one of the longest running, active ML models in healthcare showing $500K-$1M dollars in retained revenues and cost savings annually at a large health system. The model showed good discrimination in the development cohort with an AUC of 0.79 (95% confidence interval 0.79 - 0. 80) and good discrimination in the validation cohort with an AUC of 0.73 (95% confidence interval 0.72-0.73). 

Service Line

Perioperative

Published Outcomes

ScienceDirect

CLINICALTRIALS-LLM

Problem

Significant challenges exist in clinical trial operations, specifically the efficient identification, matching, and recruitment of patients, including those from underserved populations. The process is very cumbersome and rarely timely which is a problem for getting patients started in clinical trials, especially in Oncology, where time to treatment literally is life and death.

Solution

Our LLM consumes large volumes of publicly available data on clinical trials and then using our Automated Chart Review LLM (AUTOCHART), it matches the trials to patients using all inclusion and exclusion criteria. Our systematic evaluations show that our automation solution can accurately predict complex, criterion-level eligibility with faithful explanations significantly reducing staff time and resource requirements to recruit for and manage clinical trials. This allows health systems to increase the volume of clinical trials they can facilitate with their pharma and life sciences partners.

Service Line

Hospital-Wide

PACU Length of Stay predictor (PACU-LOS)

Problem

Health systems often face inefficiencies in PACU (Post-Anesthesia Care Unit) scheduling, leading to prolonged stays that strain resources and increase costs. This issue intensifies with complex cases, where the time spent in the PACU often exceeds initial estimates. Afternoon surgeries, in particular, are at increased risk of extended stays that may require unexpected inpatient admissions. These issues disrupt surgical schedules and impose significant financial burdens, including increased staffing demands and the costs of overnight care.

Solution

We have developed and validated a preoperative prediction ML model for prolonged PACU-LOS after ambulatory surgery. This model helps optimize PACU utilization, enhancing efficiency and cost-effectiveness.

Outcome

Our model demonstrated outstanding discriminatory ability across development, internal, and external validation cohorts, achieving areas under the receiver operating characteristic curve of 0.82, 0.82, and 0.80, respectively. For surgeries starting in the afternoon, PACU-USE scores of ≥43 predicted a 32% risk of PACU stays past 8 PM, compared to just 8% for lower scores (p<0.001). This resulted in a higher direct PACU cost of care of $207 (p<0.001). The benefits of using the PACU-USE score were especially significant in freestanding ASCs without PACU bed limitations.

Service Line

Perioperative

Post Operative Respiratory Failure Prediction (RESPIRATORY-POSTOP)

Problem

Postoperative respiratory failure remains a critical issue in surgical care, associated with significantly increased odds of mortality, higher healthcare costs, and diminished quality of life. Despite ongoing research and various definitions of respiratory failure by notable healthcare organizations, the incidence rate of this complication has not improved in the past two decades, hovering between 1–4%. Existing prediction tools for postoperative respiratory failure suffer from limitations such as poor performance, narrow applicability, and manual calculation requirements, which hinder their widespread implementation. These tools often fail to integrate modern machine learning techniques, which can more effectively handle complex data interactions and offer improved accuracy.

Solution

This machine learning prediction model has set a new standard for accurately identifying postoperative respiratory failure, enhancing patient outcomes and improving healthcare quality metrics.

Outcome

Our cutting-edge ML model outperforms existing tools, demonstrating superior accuracy with an AUROC of 0.93, compared to 0.82 for ARISCAT and SPORC-1. At 80-90% sensitivities, our model achieves a higher positive predictive value (11%) and a lower false positive rate (12%) compared to ARISCAT and SPORC-1.

Service Line

Respiratory

Published Outcomes

ScienceDirect

Prediction of Prolonged Ventilation (RESPIRATORY-RISK)

Problem

Acute respiratory failure occurs frequently in hospitalized patients and often starts before ICU admission. A risk stratification tool to predict mortality and risk for mechanical ventilation would enable earlier evaluation and intervention leading to improved outcomes.

Solution

We developed and validated an ML model integrated within the EHR to identify patients at risk of respiratory failure.

Outcome

The hospital mortality for patients on mechanical ventilation >= 48 h was 33% in both 2013 and 2017 compared to mortality rates of 1.4% in 2013 and < 1% in 2017 for patients with MV < 48 h.

Service Line

Respiratory

Published Outcomes

Springer

SEPSIS-LLM

Problem

The lack of Sepsis-specific LLM training data, the numerous systems which house data markers required to accurately asses Sepsis and the volume of patients on any given day in a typical hospital make predicting and optimizing the management of Sepsis very challenging.

Solution

Bundled with our Sepsis ML model, our computable phenotyping LLM uses clinical notes and other sources of data to identify severe sepsis based on large volumes of provider documentation of patient progress and clinical conditions across several systems in real-time and integrated within the clinical workflow.

Service Line

Hospital-Wide

Sepsis Realtime Phenotyping (SEPSIS-REALTIME)

Problem

Existing methods of defining severe sepsis are problematic for several reasons. They are dependent on diagnosis codes, which are only available after patient discharge, limiting timely intervention. The discovery process often relies on chart abstraction, which is labor-intensive and inefficient. Automated methods tend to be inaccurate, generating a large number of false positives; for instance, some models erroneously indicate that 25% of patients have sepsis. Additionally, no current methods effectively utilize clinical text, and the existing models are not practically useful in real-world clinical settings.

Solution

Our custom-built phenotype offers a fully automated, clinically validated solution that accurately reflects the actual prevalence of sepsis in the hospital. It can be deployed in real time and leverages clinical notes for enhanced accuracy.

Service Line

Hospital-Wide

Published Outcomes

Arxiv

Model to Predict Risk of Cancer Metastasis to the Spinal Chord (ONCOLOGY-SPINAL)

Problem

Patients who developed spinal metastesis weren't being evaluated fast enough by oncologists.

Solution

Our cutting-edge NLP model demonstrates exceptional accuracy in predicting spinal tumors and MECC in spine MRI reports. Integration within the EHR ensures faster referrals to specialists. This advancement promises to reduce morbidity and increase survival rates, significantly enhancing patient outcomes.

Outcome

Out of 37,579 radiology reports reviewed, 36,676 were labeled negative and 903 were identified with MECC. By setting a positive result cutoff at 0.02 to minimize false negatives, we achieved a 100% sensitivity rate and a remarkably low false positive rate of 2.2%.

Service Line

Oncology

Automated Concept Extraction and Encoding Services for Billing, Reporting, QI & PI

Problem

Solution

Our solution revolutionizes the automated extraction of key concepts from complex clinical notes. By leveraging a systematic and standardized approach, it surpasses current state-of-the-art extraction algorithms in both performance and generalizability.

Service Line

Quality

ExplainerAI^TM Analytics Dashboard Platform

Performance Insights, Management and Governance for AI / ML Models

Description