黑料吃瓜群网

An artificial intelligence (AI) tool to improve the safe use of intravenous heparin is being developed by researchers at (PAH).

The development has been made possible due to Queensland Health鈥檚 significant investment in implementing electronic medical records (EMRs) into 16 public hospitals over the last seven years.

The AI tool utilises machine learning coupled with data from the EMR to predict the therapeutic effect for a given dose of the drug, as measured by a blood test called the activated thromboplastin time (aPTT). This prevents underdosing which would not eliminate clots, or overdosing which could cause a bleed.

Professor Ian Scott, Director of Internal Medicine and Clinical Epidemiology at PAH, said, 鈥淎 lot of patients are admitted to hospital with conditions such as acute coronary syndrome, deep venous thrombosis and pulmonary embolism and they need to be treated with IV heparin to dissolve a blood clot or thin their blood to prevent further clots.

鈥淚V heparin allows you to achieve therapeutic effects fairly quickly and also has the advantage of a very short half-life. This means the effects will clear pretty quickly when dosing is stopped if the patient starts to bleed.

鈥淭he problem with IV heparin dosing is that individuals vary markedly in how they metabolise this drug, so each patient鈥檚 response is hard to predict.鈥�

He said aPTT testing would ideally be conducted within six to eight hours of the initial dose and then at regular intervals so that the infusion dose of heparin could be adjusted accordingly.

鈥淲hat we have found is that in many cases, it takes a long time to reach a therapeutic aPTT in patients who have potentially life-threatening conditions. In a study of about 200 patients at PA 黑料吃瓜群网, it took a median time of 36 hours to achieve a therapeutic aPTT.

鈥淭he aim of the research using the AI model is to answer the question: if we give this patient a certain bolus dose and a certain maintenance infusion, what is the aPTT likely to be within the next 12 hours?鈥� Scott said.

EMR data from 2783 hospital patient admissions across four hospitals in Metro South from 2017 to 2020* was gathered on a host of patient factors such as weight, biochemistry, full blood count, past medical history, age and sex, as well as various other clinical measures. Across all these admissions, at 12 hours after initiation of IV heparin:

• about 60% of aPTTs were below the therapeutic range;
• 17% above (toxic);
• 23% within therapeutic range.
  听

鈥淲orking with data scientists from the Clinical Informatics team here at Metro South, we developed a machine learning model, using a commercial software package, which was shown to be highly accurate in predicting which patients would have a subtherapeutic aPTT and moderately accurate in predicting which patients would have a supra-therapeutic aPTT,鈥� Scott said.

Senior Director of Digital Health and Informatics at Metro South Stephen Canaris said the strength of the collaboration between data scientists, informatics teams and the clinicians involved in the project was about bringing the value of each specialty to the development and validation of the model.

鈥淭here will always be questions that data scientists can鈥檛 answer which will require clinical knowledge to determine the choice and range of patient factors to collect, but machine learning means that a host of factors that might affect the outcomes will be teased out in the process that clinicians won鈥檛 have intuitively known about,鈥� Canaris said.

鈥淚n this instance, machine learning was able to identify about 93 factors (or variables) that were influential 鈥� but not equally important 鈥� in deciding how an individual patient responds to IV heparin. In partnership with clinicians, we were able to grade these to develop the top 10, followed by another 20 or 30 variables, that had the most predictive value,鈥� he said.

鈥淎fter developing the AI tool, we then moved to 鈥榠nternal validation鈥� where we looked at the predictive performance of the model which showed it had reasonably good accuracy,鈥� Canaris said.

Scott explained the importance of these results.

鈥淭he model was highly accurate in predicting which patients would have a sub-therapeutic range which has considerable implications for preventing underdosing of patients, particularly those who have suffered a severe pulmonary embolism,鈥� he said.

鈥淲e are now backward engineering the model so that we can answer the question clinicians will ask: what bolus dose and what maintenance dose should I prescribe in order to achieve an aPTT of say, 70 or 80 seconds within the next six to 12 hours?鈥�

Once this retrofitted model has been developed, the next steps will involve further validation studies using live data to test its accuracy, then convert it into a prototype application (app) to apply to the EMR, and, if all goes well, work towards obtaining approval from the Therapeutic Goods Administration given this is a software-as-a-medical-device application.

鈥淎n important step in proving the clinical utility of this AI tool will be a prospective clinical trial where the tool is used in one hospital (or unit) but not in another, and we compare the outcomes of patients receiving IV heparin in sites that use or do not use the tool鈥� Scott said.

鈥淎side from accuracy, the value-add for this research will be integrating the tool within the EMR alongside other models currently in development, such as an early warning alert tool for clinically deteriorating patients developed by ICU/CSIRO researchers, and other predictive models focused on patient medication safety that we are currently developing.

鈥淲e have a way to go, but it鈥檚 a promising start.鈥�

^{*The model was first trained and internally validated on data from 2783 admissions to PA, Logan, QEII and Redland 黑料吃瓜群网s, and then underwent external validation using data from 236 admissions to Gold Coast 黑料吃瓜群网.}

Originally published by Queensland Government, Metro South Health .

Image credit: iStock.com/toeytoey2530