MRIdle – from data collection to real-live intervention, to reduce appointment no-shows at the radiology ward
Wissensdatenbank Technologie C.2: Reduktion der Leerlaufzeiten von MRI-Systemen in Radiologieabteilungen durch den Einsatz von KI-basierter Terminplanungssoftware - MRIdleNo-show appointments in healthcare lead to wasted resources and poor patient outcomes. This study, using data from the University Hospital of Zurich, develops predictive models to forecast and reduce no-show rates by focusing on data quality. Accurate predictions enable targeted interventions, improving appointment adherence and optimizing healthcare efficiency. Our findings offer practical solutions to mitigate the financial and health impacts of missed appointments.
Problembeschreibung, Forschungsfrage und Relevanz
No-show appointments represent an ongoing challenge to the healthcare system.
Each missed appointment results in wasted efforts from the scheduling team, idle equipment, and underutilized medical personnel. This inefficiency leads to significant financial costs, estimated between $1673 to $1965 per no-show, and poor health outcomes due to delayed diagnoses and treatments. Previous research has indicated that no-show rates can range from 3% to 31%, highlighting the need for effective predictive models to mitigate this issue.
Our research aims to address the question: "How can data quality be leveraged to predict and reduce no-show rates in clinical settings?" By focusing on data quality, we aim to develop robust predictive models that accurately forecast no-show appointments. The study utilizes data from the University Hospital of Zurich's Radiology Information System (RIS) and electronic medical records system (KISIM).
Relevance: Addressing no-show rates is critical for optimizing resource utilization, improving patient care, and reducing healthcare costs. By developing accurate predictive models, we can identify high-risk patients and implement targeted interventions, such as reminder calls, to ensure better appointment adherence. This research not only contributes to the academic understanding of predictive modeling in healthcare but also provides practical solutions for improving operational efficiency in medical institutions. Ultimately, improving the prediction and management of no-show appointments can lead to better health outcomes by ensuring timely access to medical care, thereby reducing morbidity and mortality associated with delayed treatments.
Methoden und Vorgehen im Projekt
We collected and transformed data from the University Hospital of Zurich's Radiology Information System (RIS) and electronic medical records system (KISIM), with detailed patient information such as age, sex, occupation, marital status, and distance to the hospital.
We employed three machine learning models: Logistic Regression, Random Forest, and XGBoost. We split our dataset into training and validation sets (80%-20%), performing 5-fold cross-validation on the training set to select the best model configurations. Models were evaluated on the hold-out validation set using standard metrics and a "Top-10 Precision" metric, which reflects the practical application of predicting the top 10 most likely no-shows each week.
An additional validation step, called a ‘Silent Live Test’ (SLT), involved using the models to make live no-show predictions, which were then compared to actual outcomes. This ensured the models' robustness before deployment.
Finally, we implemented an intervention study where scheduling staff called the top 10 patients predicted to be most likely to no-show each week, reminding them of their appointments. This intervention was designed to reduce the no-show rate and optimize resource utilisation in the radiology department.
Ergebnisse und Erkenntnisse
We used the ‘audit history’ raw data from RIS to create a dataset of appointments along with their show/no-show outcome
Our models were shown to give a top-10 precision with a marked improvement over the baseline value. Meaning that, theoretically (i.e. with a well performed intervention), the appointments in this smaller group could be targeted to prevent no-shows, leading to more efficient usage of the MRI machines, and consequently better patient care.
The Silent Live Test was performed and validated our accuracy metrics, and helped us to set-up the infrastructure and planning of the real-life intervention
We performed an intervention study over a nine-month period, from March until December 2023. Each Monday and Thursday, the scheduling staff at the radiology department received a list of upcoming appointments which the model deemed at high risk of no-showing. We also collected a control group during this time – these patients were not contacted, but the outcome of their appointment was noted for comparison. The nurses then contacted
Tthe intervention patients by telephone, reminding them of their appointments. Results from this intervention will be published in due course.
Empfehlungen für die Praxis
Upcoming no-shows can be predicted with an accuracy higher than random guessing, so in theory this could be used to reduce the no-show rate at the radiology ward
We are currently deriving insights from the intervention results which will lead to concrete practice recommendations
Zitierung des Beitrags
MRIdle – from data collection to real-live intervention, to reduce appointment no-shows at the radiology ward