Skip to content

Implementing machine learning techniques to estimate the impact of underdosed DOACs, and aim patients at high bleeding risk in an elderly frail population treated for atrial fibrillation

Pdf

PDF Icon

European Statement

Clinical Pharmacy Services

Author(s)

Dorian Protzenko, Vincent Hoang, Guillaume Hache

Why was it done?

We unveiled during an audit that, in the past 2 years, 19% of our hospital DAOCs prescriptions were underdosed: due to the population profile (old, frail), the conventional bleeding risk scores were consistently high and, as such, not informative. To avoid a hypothetical bleeding risk, physicians were randomly underdosing patients beyond guidelines, without any evidence regarding the efficacy.

What was done?

Using machine learning, we unveiled that underdosing direct oral anticoagulants [DAOCs] to prevent bleeding risk in an old and frail population had no significant impact on drug-related hospitalization [DRH] nor death, and cannot be supported. To help targeting patients for whom extra care would be more beneficial rather than underdosed DAOCs, we built a predictive model of bleeding events and provided risk factors among our population.

How was it done?

We performed a retrospective study, based on data collected during the audit, of patients treated between October 2020 and April 2022 with Apixaban or Rivaroxaban for atrial fibrillation [AF]. Demographic and clinical criterias (i.e., GFR, polypathology, co-medications, prescribed DAOC, respecting dosage and scheduling) were collected. The occurrence of specific outcomes (i.e., bleeding and thrombosis that led to medical care and drug seizure, DRH and death) were retrieved from the patients’ medical records. Machine learning explorations were performed using RStudio®.

What has been achieved?

119 patients were included. We modeled using logistic regression the impact on selected outcomes of underdosing DAOCs. We found out that underdosed DAOCs were associated with a lower bleeding risk (OR=0.30, CI95%[0.07;0.95]), a higher thrombosis risk (OR=6.67, CI95%[1.23;50.0]), but without any impact on DRH nor death. Unsupervised algorithms unveiled that DAOC choice (Rivaroxaban: OR=2.80, CI95%[1.15;7.13]), sex (Male: OR=0.44, IC95%[0.16;1.12]) and using dosages from guidelines (OR=3.32, CI95% [1.05;14.80]) were predominant explanatory variables regarding bleeding risk. The choice of DAOC was the only covariate that impacted DRH (Rivaroxaban: OR=2.78, CI95%[1.22;6.56]). Finally, using a gradient-boosting algorithm, bleeding risk was predicted with a 0.73 roc-auc, superior to conventional models.

What next?

Therapeutic education of patients and caregivers, telephone follow-up or pharmaceutical consultations will be implanted for patients at high bleeding risk. An audit will be performed next year to measure underdosed prescriptions rate, and improve the model with new data.

×

Deadline extended to July 15th

Problems caused by shortages are serious, threaten patient care and require urgent action.

Help us provide an overview of the scale of the problem, as well as insights into the impact on overall patient care.

Our aim is to investigate the causes of medicine and medical device shortages in the hospital setting,  while also gathering effective solutions and best practices implemented at local, regional, and national levels.

×

Join us in Prague for the 2nd edition of BOOST!

Secure your spot in the Movement for Shortage-Free World

BOOST is where visionaries, innovators, and healthcare leaders come together to tackle one of the biggest challenges in hospital pharmacy—medicine shortages.