Supporting clinical professionals in the decision-making for patients with chronic diseases Mitja Luštrek 1, Božidara Cvetković 1, Maurizio Bordone 2, Eduardo Soudah 2, Carlos Cavero 3, Juan Mario Rodríguez 3, Aitor Moreno 4, Alexander Brasaola 4, Paolo Emilio Puddu 5 1 Jožef Stefan Institute, Slovenia 2 CIMNE, Spain 3 Atos, Spain 4 Ibermática, Spain 5 University of Rome “La Sapienza”, Italy
Rationale Medical labs produce a lot of data on a patient Telemonitoring produces even more data The amount of medical literature is huge Overwhelming for a clinical professional
Rationale Medical labs produce a lot of data on a patient Telemonitoring produces even more data The amount of medical literature is huge Overwhelming for a clinical professional Needs tools to make sense of all these data Decision support system (DSS)
Clinical workflow 1.The doctor starts examining the condition of a patient, possibly because of an alert by the DSS.
Clinical workflow 1.The doctor starts examining the condition of a patient, possibly because of an alert by the DSS. 2.The doctor examines the patient’s current (and historic) risk, computed by the DSS.
Clinical workflow 1.The doctor starts examining the condition of a patient, possibly because of an alert by the DSS. 2.The doctor examines the patient’s current (and historic) risk, computed by the DSS. 3.If the risk is high, the doctor looks for reasons. The DSS computes the contribution to the risk for each of the monitored parameters.
Clinical workflow 1.The doctor starts examining the condition of a patient, possibly because of an alert by the DSS. 2.The doctor examines the patient’s current (and historic) risk, computed by the DSS. 3.If the risk is high, the doctor looks for reasons. The DSS computes the contribution to the risk for each of the monitored parameters. 4.The doctor may look for further information in the medical literature with the help of the DSS.
Clinical workflow 1.The doctor starts examining the condition of a patient, possibly because of an alert by the DSS. 2.The doctor examines the patient’s current (and historic) risk, computed by the DSS. 3.If the risk is high, the doctor looks for reasons. The DSS computes the contribution to the risk for each of the monitored parameters. 4.The doctor may look for further information in the medical literature with the help of the DSS. 5.The doctor may reconfigure the DSS.
DSS architecture Electronic health record Sensors Literature consultation External data Risk assessment Expert knowledge Machine learning Anomaly detection Alerts Configuration
Risk assessment – expert knowledge Electronic health record Sensors Literature consultation External data Risk assessment Expert knowledge Machine learning Anomaly detection Alerts Configuration
Monitored parameters Search of medical literature for parameters affecting the risk (for congestive heart failure) Survey among 32 cardiologists to determine the importance of these parameters
Monitored parameters Search of medical literature for parameters affecting the risk (for congestive heart failure) Survey among 32 cardiologists to determine the importance of these parameters Additional information for each parameter: – Minimum, maximum value – Whether larger value means higher or lower risk – Values indicating green, yellow or red condition – Frequency of measurement (low = static, medium = measured by the doctor, high = telemonitored)
Risk assessment models Normalize parameter values: [0, 1] interval, 0 = lowest risk, 1 = highest risk
Risk assessment models Normalize parameter values: [0, 1] interval, 0 = lowest risk, 1 = highest risk Long-term model: sum of normalized values, weighted by their importance Medium-term model: low-frequency parameters weighted by 1/3 Short-term model: low-frequency parameters weighted by 1/9, medium-term by 1/3
Prototype
Risk assessment – machine learning Electronic health record Sensors Literature consultation External data Risk assessment Expert knowledge Machine learning Anomaly detection Alerts Configuration
Risk assessment – machine learning 1.Training data: [parameter values, cardiac event or no event] 2.Feature selection, decorrelation 3.Machine learning model selection: multilayer perceptron with input (parameters), hidden, and output (risk) layer 4.Training: 85 % accuracy on a public heart disease dataset
Risk assessment – anomaly detection Electronic health record Sensors Literature consultation External data Risk assessment Expert knowledge Machine learning Anomaly detection Alerts Configuration
Risk assessment – anomaly detection Detect anomalous (= not observed before) parameter values and their relations +No knowledge or data labeled with cardiac events needed –Anomalies do not alway mean higher risk
Risk assessment – anomaly detection Detect anomalous (= not observed before) parameter values and their relations +No knowledge or data labeled with cardiac events needed –Anomalies do not alway mean higher risk More on this in a separate presentation in this session by Božidara Cvetković
Literature consultation Electronic health record Sensors Literature consultation External data Risk assessment Expert knowledge Machine learning Anomaly detection Alerts Configuration
Literature consultation Free text / PICO question Query Free text (EHR) contextualization Ontology maping Semantic search Resources: PubMed Cochrane Library... Results Annotate, evaluate Ranking
Literature consultation Free text / PICO question Query Free text (EHR) contextualization Ontology maping Semantic search Resources: PubMed Cochrane Library... Results Annotate, evaluate Ranking
Literature consultation Free text / PICO question Query Free text (EHR) contextualization Ontology maping Semantic search Resources: PubMed Cochrane Library... Results Annotate, evaluate Ranking
Literature consultation Free text / PICO question Query Free text (EHR) contextualization Ontology maping Semantic search Resources: PubMed Cochrane Library... Results Annotate, evaluate Ranking
Literature consultation Free text / PICO question Query Free text (EHR) contextualization Ontology maping Semantic search Resources: PubMed Cochrane Library... Results Annotate, evaluate Ranking
Literature consultation Free text / PICO question Query Free text (EHR) contextualization Ontology maping Semantic search Resources: PubMed Cochrane Library... Results Annotate, evaluate Ranking
Literature consultation Free text / PICO question Query Free text (EHR) contextualization Ontology maping Semantic search Resources: PubMed Cochrane Library... Results Annotate, evaluate Ranking
Alerts and configuration Electronic health record Sensors Literature consultation External data Risk assessment Expert knowledge Machine learning Anomaly detection Alerts Configuration
Alerts and configuration Alerts: Rule engine using the Drools platform Rules triggered on parameter or risk values Alert modes (SMS, ) depend on the trigger
Alerts and configuration Alerts: Rule engine using the Drools platform Rules triggered on parameter or risk values Alert modes (SMS, ) depend on the trigger Configuration: Parameters to be monitored for each patient Parameter values indicating green, yellow or red condition for each patient
Conclusion DSS tailored to a (fairly generic) clinical workflow Can be used for all diseases to which the workflow is applicable Congestive heart failure as a case study
Conclusion DSS tailored to a (fairly generic) clinical workflow Can be used for all diseases to which the workflow is applicable Congestive heart failure as a case study Observational study with 100 patients starting shortly Tuning and testing once the data from the study is available