Which patients are at risk of developing symptom diagnoses that persist for more than a year in primary care? Development and external validation of a prediction model

Objectives

To train, test and externally validate a prediction model that supports General Practitioners (GPs) in early identification of patients at risk of developing symptom diagnoses that persist for more than a year.

Methods

We retrospectively collected and selected all patients having episodes of symptom diagnoses during the period 2008 and 2021 from the Family Medicine Network (FaMe-Net) database. From this group, we identified symptom diagnoses that last for less than a year and symptom diagnoses that persist for more than a year. Multivariable logistic regression analysis using a backward selection was used to assess which factors were most predictive for developing symptom diagnoses that persist for more than a year. Performance of the model was assessed using calibration and discrimination (AUC) measures. External validation was tested using data between 2018 and 2022 from AHON-registry, a primary care electronic health records data registry including 73 general practices from the north and east regions of the Netherlands and about 460, 795 patients.

Results

From the included 47,870 patients with a symptom diagnosis in the FaMe-Net registry, 12,481 (26.1%) had a symptom diagnosis that persisted for more than a year. Older age (≥ 75 years: OR = 1.30, 95% CI [1.19, 1.42]), having more previous symptom diagnoses (≥ 3: 1.11, [1.05, 1.17]) and more contacts with the GP over the last 2 years (≥ 10 contacts: 5.32, [4.80, 5.89]) were predictive of symptom diagnoses that persist for more than a year with a marginally acceptable discrimination (AUC 0.70, 95% CI [0.69–0.70]). The external validation showed poor performance with an AUC of 0.64 ([0.63–0.64]).

Conclusion

A clinical prediction model based on age, number of previous symptom diagnoses and contacts might help the GP to early identify patients developing symptom diagnoses that persist for more than a year. However, the performance of the original model is limited. Hence, the model is not yet ready for a large-scale implementation.