Invented by Carmelo Velez, Timothy Tschampel, Emilia Apostolova, Adam Boris, Computer Technology Associates Inc
The Computer Technology Associates Inc invention works as follows
A disease-specific ontology, crafted by expert clinicians in consensus, can be used to provide semantic meaning/characterize dynamically changing EMR data for patients in critical care settings. Concept maps (Vmaps), which are directed node-edge graphs, developed using a graphical editor and an end-user-friendly graphical interface, can be used as a first step to build disease-specific ontologies. Vmaps for disease domains that reflect expert clinical reasoning in relation to acute illnesses seen in critical care settings. ICUs that extend core clinical ontologies, developed by experts and reviewed, are then extended with existing medical terminologies and translated automatically to a domain-ontology processing engine. Semantically-enhanced EMR data derived from the ontology processing engine is incorporated into both real-time ?track and trigger? Using aggregated data, we can create rule engines and machine-learning training algorithms. The resultant rule engines and machine learning models provide enhanced diagnostic information and prognostic data respectively to assist with clinical dual modes (analytical models and models based upon experiential data) in acute critical care settings.Background for A disease-specific ontology guided rule engine and machine intelligence for enhanced critical care decisions
Automated rule-based Clinical Decision Support (?CDS)” The use of electronic medical records (?EMR?) Data has shown to improve clinician situational awareness within critical care settings. CDS tools are generally designed to “cast a broad net” and generate alerts which are not specific. Alerts generated by such tools are often ignored. Some rule-based CDS tools increase their specificity by raising the clinical criteria risk thresholds. However, in real clinical practice these augmented rule based tools may be “insufficiently sensitive” and miss important, less obvious, cases that are not recognized in early stages.
The current CDS tools only apply a relatively simple rule logic to observed data. When vital signs or laboratory results exceed preset thresholds. These rule-based CDS instruments fail to reflect complex clinical reasoning that expert clinicians use to make patient-specific decision based upon a combination of observed signatures and associated trends in potentially multiple observed data elements, along with the patient’s context and recent evidence/guidelines. Most importantly, they do not reflect expert clinician experience gained over years of practice. This results in a phenomenon called “alert fatigue”. Alert fatigue, or the high rate of overriding alerts, is a problem that has become more and more prevalent in medicine. Doctors and other healthcare providers seem to ignore computerized alerts generated by current EMR systems in between 49% to 96%, including clinically significant alerts. Medical errors are frequent in critical care and can cause mortality and morbidity in critically ill patients without effective CDS tools. Recent calls for improving CDS to support diagnostic reasoning have increased in popularity. “In summary, current CDS tools, which are based on rudimentary reasoning, found in EMRs that are commercially available, do not improve patient outcomes, such as reducing mortality in critically ill patients in hospitals.
Researchers have found that clinicians use two types of reasoning to make clinical decisions. The first is experiential and can be described as tacit, fast, intuitive, recognition-primed, implicit, acquired through exposure. The second model, rule-based, is described as rational, deliberate, slower, analytic, explicit, and conscious. CDS tools can support both modes of reasoning by using expert-trained rule engines and supervised machine-learning techniques. To recognize clinical patterns of experience in Electronic Medical Record (EMR) “Real-time data streams based on historical clinical data repositories.
Using abstracted cohort EMR data, ML can be used as a predictive model to identify patterns in nonspecific clinical data (e.g. Physiological time series associated with an outcome, e.g. Death, life-threatening illnesses, prolonged length of stay, etc.) could be associated with an outcome (e.g. The ability of ML techniques to detect/recognize subtle signatures of incipient illness probabilistically associated with diverse physiological data hours before clinical staff can detect them is an important, promising advantage for ML-based CDS over rule-based CDS. “ML-based CDS has a promising advantage over rule-based methods in that ML can detect/recognize subtle signs of illness hours before clinical staff are able to detect them.
The noisy EMR data used in CDS training can affect the accuracy of these techniques. This can lead to a degraded ML and unintended outcomes. For instance, acute syndromic conditions like sepsis and acute respiratory distress syndrome are often accompanied by comorbid conditions. For example, for acute syndromic diseases like sepsis or acute respiratory distress syndrome (?ARDS?) Outcome classification codes (e.g. ICD-9/10 is often incorrect or missing from critical care EMR records. Machine learning techniques that use EMR data can be reliant on accurate diagnostic coding. For supervised training. Label noise in ML models that are derived from high levels of training data (e.g. More than half of patients with severe sepsis do not have a documented diagnosis from their doctors. This may negatively impact ML’s predictive performance and lead to unintended consequences. Similar to missing physiological data, context can be mischaracterized, or the lack of context can result in “feature noise”. (e.g. Training data that does not distinguish between observed abnormal physiological EMR EMR training data features and “acute”? Training data that fails to distinguish observed abnormal physiological EMR training data features as?acute? Treatments or?normal responses? EMR data can be difficult to model and represent for machine learning because of its high dimensionality and dynamic nature, as well as due to noise, heterogeneity and sparsity. It also contains random errors, systematic biases, and lacks semantic harmony. The same clinical phenotype may be expressed in different ways, for example: a patient with “type 2 diabetes” can have a variety of codes and terminology. The presence of ICD-9 code 250.00, laboratory results of hemoglobin A1C higher than 7.0 or a diagnosis of ‘type 2 diabetes mellitus’ in the free-text clinical notes can all be used to identify a patient. Mentioned in the free text clinical notes. This makes it difficult for machine-learning methods to identify patterns which produce enough precise clinical predictive models for real applications. Recent studies have shown that models that are trained without any raw data enhancements to address issues such as missing context, concept harmonization, or label/feature noise can produce technically valid, but dangerously misinformed machine learning prognostic model.
Many popular ML tools, such as neural networks, have “hidden layers”. “Many popular ML tools, such as neural networks with ‘hidden layers? The outputs of EMR data are not backed up by any rationale. With tools like neural networks, unlike rule-based systems, where the logic behind a rule is easily understandable, there is not a logical flow a person can follow. Predictions from these tools might not be helpful in critical care situations that are life-threatening. The predictive ML models that are derived from EMR data must be updated regularly to reflect current practice/protocols, and demographics of patients specific to each institution. Current retraining/recalibration of predictive models based on changing practice/protocols or institution-specific patient populations is generally labor intensive.
Therefore, ML-based CDS tools have not been widely adopted in bedside workflows.
One or more embodiments of this invention provide benefits and/or resolve one or more problems in the field with automated clinical decision-support tools based on ML-based prediction models used in critical health care. The invention reveals problem-specific, contextually rich ontological models which model the clinical reasoning and decision-making of expert critical care clinicians in order to improve the performance of automated CDS for the analysis of EMR patient narrative and physiological data. This will lead to enhanced early detection and treatment of patients at risk. These models expose the expert interpretations and semantic characterizations of features, which are then used to train ML algorithms. This allows bedside clinicians to accurately recognize and explore the implications of recorded EMR structured and free-text feature data and observed physiological time series. Automation of the “pipeline” can also be achieved. By automating the ‘pipeline? “Updates to predictive models are enabled.
The preferred embodiment uses highly intuitive graphical VFusion concepts maps (herein referred to as ‘Vmaps?).” The Vmaps are developed with an ontology editor, which is a user interface, that captures the structured reasoning of multidisciplinary experts who collaborate to interpret physiological data of patients. These expert clinicians typically extend evidence-based guidelines and reflect deep expertise gained from years of management experience. The present disclosure discloses a Vmap Editor that can be used graphically to view and describe editing changes to an ontology. edit ontological class structures, properties/relationships or rules) and/or create a new ontology These intuitive graphical expressions form the foundations of a comprehensive customized ontology that can be used by semantic reasoners, rule engines and machine learning for disease-specific CDS applications. Specifically, Vmaps graphically describe clinical reasoning that is used to derive a comprehensive disease specific computable ontology that can subsequently be used with raw EMR data to semantically identify cohorts and characterize/interpret clinical features that support the dual rule-based/experiential reasoning processes of physicians, influencing production rules, natural language processing (NLP), and the machine learning discovery process with the deep expertise, heuristics, and knowledge of subject matter experts.
DESCRIPTION of FIGURES
To describe how the above-recited advantages and features can be obtained by one or more embodiments, a detailed description will be given in reference to the specific embodiments that are shown in the accompanying drawings. These drawings are only representative of typical embodiments, and therefore do not limit the scope. The accompanying drawings will provide additional detail and specificity to the description of various embodiments.
FIG. “FIG.
FIG. “FIG.
FIG. “FIG.
FIG. “FIG.
FIG. “FIG.
FIG. “FIG.
FIG. “FIG.
FIG. “FIG.
FIG. “FIG. “An embodiment of the current disclosure compares training Lasso with semantically enhanced features and labels to training Lasso with traditional EMR raw feature and coded outcome data (ICD-9).
FIGS. “FIGS.
Click here to view the patent on Google Patents.