Mobiele menu

Development of a diagnostic prediction tool for general practice, based on routine registration data (PREDICT)

Projectomschrijving

Ontwikkeling van een diagnostische predictie-tool voor de huisarts (PREDICT)

Patiënten bezoeken hun huisarts meestal met vragen over hun gezondheid. De huisarts luistert naar de patiënt en stelt een waarschijnlijkheidsdiagnose die gebaseerd is op het verhaal, het lichamelijk onderzoek, en soms ook op basis van aanvullend onderzoek. Enkel de klacht voorspelt in belangrijke mate de kans op een juiste diagnose. Zo zal de diagnose bij een patiënt die hoest meestal een verkoudheid zijn, soms een longontsteking en zelden longkanker. Allerlei factoren zoals leeftijd, geslacht en ook het eerder hebben gehad van een longontsteking beïnvloeden die kans.

Doel en werkwijze

In dit onderzoeksproject ontwikkelen we een ondersteuningstool voor huisartsen, waarmee deze tijdens het spreekuur realtime een betere inschatting kunnen maken van de kans op een bepaalde diagnose. We gaan daarbij uit van de klacht van de patiënt en een aantal persoonlijke en contextfactoren. De ondersteuningstool geven we vorm in nauwe samenwerking met huisartsen en testen we vervolgens uit in een aantal praktijken.

Verslagen


Samenvatting van de aanvraag

This research project is a development study that will result in a diagnostic prediction tool for the general practitioner (GP). In two steps, we will develop a sophisticated model that predicts the probability of an end diagnosis based on the reason for encounter during regular GP office hours (WP 1 and 2). This model will be built into a diagnostic prediction tool that is real-time available for GPs during everyday consultations (WP 3). Finally, this tool will be tested for feasibility in daily practice (WP 4). The diagnostic process in general practice is complex. GPs often use estimated probabilities of diagnoses for their policy on additional testing and referral. The used probabilities are mostly unknown, implicit and not precise. Our research will support GPs with an evidence based tool in this diagnostic process. The model departs from the symptoms that patients present to their GP. Symptoms are perceived important data in medical history taking, but they are only seldom used as key variables in scientific research since these data are not systematically coded in practice and therefore are seldom analyzed for research. Based on the data from the practice based registration network FaMe-net, in which GPs register the reason for encounter (RFE) as a coded variable for every episode in the EHR, we will build a new diagnostic prediction model based on reasons for encounter for new episodes of care (as presented in the beginning of the consultation) and on individual patient data available from the electronic medical record (EHR) such as age, sex, socio-economic status, duration of symptoms and previous diagnoses. FaMe-net originates from the primary care practice-based research network (PBRN) affiliated to the Radboud University Medical Center. FaMe-net started collecting data in 2005. Data include 308,000 patient years and over 2.2 million encounters, and has, over the years, proven to generate reliable and valid data. In Work Package (WP) 1 we will identify – for each of the ten most prevalent RFEs - the most important end diagnoses, specified for different age and sex groups of patients presenting to their GP. This will result in straight and basic prediction models for RFE’s and their final diagnoses. At this stage, we will also experiment with free text words in the registration data and its relation to the structured and coded RFE data. This will allow us to prepare the prediction model to be used in the broader context of other EHR systems. In WP 2, a more sophisticated diagnostic prediction model of end diagnoses will be made for patients presenting with one of the ‘Top10 RFE’s, based on the model in WP 1. Besides age and sex, also other variables from the EHR will be included in the model. This will modify the probability of end diagnoses calculated in WP 1. Using a machine-learning approach based on Bayesian networks, the influence of the following variables will be investigated: duration of symptoms until first consultation within episode, all previous diagnoses, additional RFEs, contact frequency, seasonal influence, and contextual information such as ethnicity, working hours, family history and intoxications. This will result in more precise and personalized prediction models for final diagnoses of all of the Top10 RFEs based on the statistical models with best fit for each RFE. In WP 3, we will develop and build a digital application which will be incorporated in the EHR. Initially, we will focus on the EHR that is used by FaMe-Net (TransHis). This built-in application will show probabilities of final diagnoses for the top-10 RFEs that have been developed in WP1 and WP2. The tool will be prompted real-time in the EHR when the RFE is registered during a GP office visit. Besides of the RFE, the shown proportions thus will be based on all the relevant and available individual patient data registered in the EHR: age, sex, previous diagnoses and all other variables that are (possibly) contributing to the probability of an end diagnosis. In an iterative process with potential end users (GPs and GP trainees), an alfa-version of the tool will be developed into a beta-version. This will include testing of preliminary versions in daily practice. Finally, in WP 4 we will perform a broader feasibility study within FaMe-net, using qualitative methodologies to explore and study the acceptability of the prediction tool, its user-friendliness, and the impact on GPs’ clinical work and shared decision making in the consultation room. This will result in a gamma-version of the tool (release candidate), which is fit for use in daily practice, and which can serve as a candidate to be introduced in other EHR systems as well.

Kenmerken

Projectnummer:
839150005
Looptijd: 43%
Looptijd: 43 %
2021
2027
Gerelateerde subsidieronde:
Projectleider en penvoerder:
prof. dr. H.J. Schers MD PhD
Verantwoordelijke organisatie:
Radboud Universitair Medisch Centrum
Afbeelding

Onderzoek naar huisartsgeneeskunde door aioto’s

Om de behandeling van patiënten te verbeteren, financieren we onderzoek naar wetenschappelijke vragen uit de dagelijkse praktijk. Doordat de onderzoeken worden uitgevoerd door artsen in opleiding tot onderzoeker dragen ze ook bij aan de academisering van de opleidingen. Lees meer over deze onderzoeken.