Predicting peripheral venous catheter related phlebitis using machine learning (PPML): development and prospective validation of PPML for cardiology inpatients

Abstract

Background

Peripheral venous catheter use is a common healthcare practice and carries risk for peripheral venous catheter-related phlebitis (PVCP). The aims of this study were to develop a machine learning model using inpatient hospital data to accurately predict the risk of PVCP and apply this model for early identification to reduce the risk of PVCP in the department of Cardiology.

Methods

A prediction model was developed to estimate the risk of developing PVCP within 3–24 h after introduction during a clinical admission in the department of Cardiology. Data of 107.419 generic hospital clinical admissions between January 2017 and December 2020 were used. For evaluating generalizability of the model, 1.199 clinical admissions between January 2021 and May 2021 from Cardiology were used as validation dataset. For prospectively evaluating clinical utility, 9.885 admissions between May 2021 and December 2022 of Cardiology were used.

Results

Our results demonstrate a strong-performing model with an AUC of 0.89 (CI: 0.87–0.91) based on the test set, and an AUC of 0.72 (CI: 0.66–0.78) based on the validation data. A significant descending trend in the incidence of PVCP in long stay admissions (p-value of 0. 01) was observed during prospective evaluation.

Conclusions

The early identification proves beneficial in reducing the risk of PVCP significantly for patients with long-stay admissions in the Cardiology department.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12911-025-03158-6.

Introduction

Peripheral venous catheter (PVC) use is common healthcare practice, particularly in acute care settings to allow for fast drug and fluid delivery. However, it is also associated with a risk of peripheral venous catheter-related bloodstream infections (PVC-BSIs). PVC-BSIs are associated with increased morbidity, prolonged hospital stays, increased healthcare costs, and even death [1, 2]. Despite the availability of evidence-based guidelines for preventing PVC-BSIs, such as the use of chlorhexidine skin antisepsis and maximum barrier precautions during catheter insertion, implementation of these measures remains a challenge in clinical practice. During a clinical admission, most patients are given a peripheral intravenous (IV) canula either to administer medication or as part of a vascular access for high-risk rhythm monitoring. IV access exposes the patient to the risk of developing a PVC-BSI, and possibly suffer from complications such as phlebitis and thrombophlebitis. These two complications are the most prevalent within PVC-BSIs and have a reported incidence that can be as high as 19% [3]. When left unattended, peripheral venous catheter-related phlebitis (PVCP) poses a considerable risk potentially escalating to more severe complications. Early intervention for patients at risk is therefore pivotal, and although evidence-based guidelines are in place for prevention, studies from a survey show that adherence to these was less than 50% among healthcare facilities. These findings highlight the need for innovative solutions to improve implementation of preventive measures and reduce the burden of PVCP [4, 5].

In recent years, machine learning algorithms have emerged as powerful tools in healthcare settings, offering the potential to revolutionize infection control strategies. Recent studies have demonstrated the effectiveness of risk prediction models across various clinical settings, highlighting their potential to support early intervention and improve patient outcomes [6–8]. By leveraging large-scale patient data, including vital signs, laboratory results, patient history, and other relevant parameters, machine learning models can detect subtle patterns and changes that may precede the onset of PVCP. Several studies show that machine learning algorithms (such as RandomForest or XGBoost) perform adequately in early detection or identification of patients who are at risk for PVCP. This may provide crucial information for implementing prompt interventions to prevent more severe complications [9–12]. To the best of our knowledge, studies applying a machine learning model to aid early identification of phlebitis in a clinical setting are scarce. Furthermore, the clinical utility over time when such a model would be implemented into the workflow of health care professionals in a real-world hospital department remains unclear [13].

The objectives of this study were twofold: first, to assess the performance of a machine learning-based model capable of predicting PVCP within every 3–24 h during clinical admission; and the second objective was to evaluate the practicality of applying this model as a prediction tool over a minimum period of 1 year in a real-world hospital department. The study focused on the Cardiology department as a case study for implementation and evaluation.

Our hypotheses were as follows: (1) a robust and broadly applicable machine learning model developed on hospital data can accurately predict the risk of developing PVCP; and (2) applying this model as a prediction tool for early identification would enable timely interventions and reduce the risk of developing PVCP.

Methods

Study design

For this study, a prediction model based on machine learning algorithms was developed to estimate the risk of PVCP within every 3–24 h after IV introduction during a clinical admission in the department of Cardiology. PVCP was defined by localized symptoms such as pain, redness, swelling, and warmth along the vein where the catheter is placed. PVCP occurrences were extracted using a rule-based SQL query searching through clinical notes. The rule-based SQL query was based on an expansive list of keywords, also filtering out negations, returning a dataset of clinical admissions related to the extracted clinical notes. This dataset was thoroughly manually validated through in two rounds of validation to ensure a good data quality, and more specific information of the list of keywords can be found in the supplementary materials.

Setting and population

To develop the prediction model, 107.419 clinical admissions of all hospital departments between January 2017 and December 2020 in total were included, with 2.170 cases (∼ 2%) which developed a PVCP. To validate the performance of the model specifically for the department of Cardiology, clinical admissions (acute and long stay) between January 2021 and May 2021 were used with a total of 1.199 clinical admissions with 106 cases (∼ 9%) of PVCP during admission. To evaluate clinical utility of the model after go-live, clinical admissions between May 2021 and December 2022 were included with 9.885 clinical admissions from the department Cardiology, and 352 (∼ 4%) of PVCP during admission. Cardiology as department was included for this validation due to the high incidence of PVCP as compared to other hospital departments. For more information concerning the population see Table 1. The use of patient data for this study was approved and registered with the institutional review board (METC).

Table 1.

Patient characteristics of the retrospective hospital data (Jan 2017 – Dec 2020)

	PVCP		P-value	Total (N = 107419)
	Present (N = 2170)	Absent (N = 105249)
Age
Mean (SD)	66.3 (15.9)	50.7 (19.8)	< 0.001	51.0 (19.9)
Median [Min, Max]	70.0 [18.0, 99.0]	49.0 [18.0, 119]		50.0 [18.0, 119]
Sex
Female	953 (43.9%)	61,884 (58.8%)	< 0.001	62,837 (58.5%)
Male	1217 (56.1%)	43,365 (41.2%)		44,582 (41.5%)
Body Mass Index
Mean (SD)	26.3 (5.13)	26.1 (5.01)	0.0834	26.1 (5.01)
Median [Min, Max]	25.8 [12.3, 40.8]	25.4 [12.1, 40.8]		25.4 [12.1, 40.8]
History of smoking
No	1816 (83.7%)	90,916 (86.4%)	< 0.001	92,732 (86.3%)
Yes	354 (16.3%)	14,333 (13.6%)		14,687 (13.7%)
Use of alcohol
No	1569 (72.3%)	76,688 (72.9%)	0.579	78,257 (72.9%)
Yes	601 (27.7%)	28,561 (27.1%)		29,162 (27.1%)
Systolic blood pressure
Mean (SD)	132 (24.0)	127 (20.1)	< 0.001	127 (20.2)
Median [Min, Max]	133 [69.0, 184]	125 [69.0, 184]		125 [69.0, 184]
Diastolic blood pressure
Mean (SD)	77.1 (10.9)	73.8 (11.7)	< 0.001	73.8 (11.7)
Median [Min, Max]	77.0 [40.0, 107]	74.0 [40.0, 107]		74.0 [40.0, 107]
Temperature (Celsius)
Mean (SD)	37.1 (0.476)	36.9 (0.512)	< 0.001	36.9 (0.512)
Median [Min, Max]	37.1 [35.6, 38.3]	36.9 [35.6, 38.3]		36.9 [35.6, 38.3]
Platelet count
Mean (SD)	284 (103)	256 (86.2)	< 0.001	256 (86.7)
Median [Min, Max]	280 [23.0, 498]	248 [23.0, 498]		249 [23.0, 498]
Creatinine blood
Mean (SD)	77.7 (21.3)	71.6 (19.7)	< 0.001	71.7 (19.8)
Median [Min, Max]	77.0 [15.1, 137]	69.0 [13.5, 137]		69.0 [13.5, 137]
Glucose level
Mean (SD)	7.92 (1.39)	6.99 (1.55)	< 0.001	7.01 (1.55)
Median [Min, Max]	8.00 [2.80, 12.5]	6.80 [2.30, 12.5]		6.80 [2.30, 12.5]
Leukocytes
Mean (SD)	9.61 (3.95)	11.1 (5.08)	< 0.001	11.1 (5.07)
Median [Min, Max]	8.80 [0.500, 26.5]	9.50 [0, 26.7]		9.50 [0, 26.7]
Count of active medication
Mean (SD)	13.0 (10.2)	4.77 (5.60)	< 0.001	4.94 (5.84)
Median [Min, Max]	11.0 [0, 98.0]	3.00 [0, 73.0]		3.00 [0, 98.0]
Use of anticoagulants
No	143 (6.6%)	51,378 (48.8%)	< 0.001	51,521 (48.0%)
Yes	2027 (93.4%)	53,871 (51.2%)		55,898 (52.0%)
Use of nonsteroidal anti-inflammatory drugs
No	1591 (73.3%)	58,417 (55.5%)	< 0.001	60,008 (55.9%)
Yes	579 (26.7%)	46,832 (44.5%)		47,411 (44.1%)
Use of corticosteroids
No	1570 (72.4%)	90,823 (86.3%)	< 0.001	92,393 (86.0%)
Yes	600 (27.6%)	14,426 (13.7%)		15,026 (14.0%)
Clinical admissions last year
Mean (SD)	0.159 (0.600)	0.0681 (0.353)	< 0.001	0.0699 (0.360)
Median [Min, Max]	0 [0, 9.00]	0 [0, 9.00]		0 [0, 9.00]
Emergency Department visits last 6 months
Mean (SD)	0.0963 (0.449)	0.0509 (0.291)	< 0.001	0.0518 (0.296)
Median [Min, Max]	0 [0, 8.00]	0 [0, 11.0]		0 [0, 11.0]
Length of Stay
Mean (SD)	11.8 (17.6)	2.75 (6.27)	< 0.001	2.93 (6.81)
Median [Min, Max]	7.00 [0, 492]	1.00 [0, 368]		1.00 [0, 492]
Hypercholesteremia
No	1917 (88.3%)	98,448 (93.5%)	< 0.001	100,365 (93.4%)
Yes	253 (11.7%)	6801 (6.5%)		7054 (6.6%)
Diabetes
No	1626 (74.9%)	92,933 (88.3%)	< 0.001	94,559 (88.0%)
Yes	544 (25.1%)	12,316 (11.7%)		12,860 (12.0%)
Hypertension
No	1300 (59.9%)	83,624 (79.5%)	< 0.001	84,924 (79.1%)
Yes	870 (40.1%)	21,625 (20.5%)		22,495 (20.9%)
Atrial fibrillation
No	1821 (83.9%)	99,961 (95.0%)	< 0.001	101,782 (94.8%)
Yes	349 (16.1%)	5288 (5.0%)		5637 (5.2%)
Renal insufficiency
No	1730 (79.7%)	98,971 (94.0%)	< 0.001	100,701 (93.7%)
Yes	440 (20.3%)	6278 (6.0%)		6718 (6.3%)
Cerebrovascular disorders
No	2007 (92.5%)	102,257 (97.2%)	< 0.001	104,264 (97.1%)
Yes	163 (7.5%)	2992 (2.8%)		3155 (2.9%)
Ischemic cardiovascular disease
No	1798 (82.9%)	98,848 (93.9%)	< 0.001	100,646 (93.7%)
Yes	372 (17.1%)	6401 (6.1%)		6773 (6.3%)
Peripheral vascular disease
No	2139 (98.6%)	104,778 (99.6%)	< 0.001	106,917 (99.5%)
Yes	31 (1.4%)	471 (0.4%)		502 (0.5%)
Heart failure
No	1804 (83.1%)	101,965 (96.9%)	< 0.001	103,769 (96.6%)
Yes	366 (16.9%)	3284 (3.1%)		3650 (3.4%)
Cardiovascular disease
No	2151 (99.1%)	104,824 (99.6%)	0.00127	106,975 (99.6%)
Yes	19 (0.9%)	425 (0.4%)		444 (0.4%)
Skin disease
No	1805 (83.2%)	95,237 (90.5%)	< 0.001	97,042 (90.3%)
Yes	365 (16.8%)	10,012 (9.5%)		10,377 (9.7%)
Count of intravenous medication
Mean (SD)	9.15 (8.32)	5.10 (4.34)	< 0.001	5.19 (4.49)
Median [Min, Max]	7.00 [0, 97.0]	4.00 [0, 106]		4.00 [0, 106]
Cardiovascular medication
No	372 (17.1%)	59,064 (56.1%)	< 0.001	59,436 (55.3%)
Yes	1798 (82.9%)	46,185 (43.9%)		47,983 (44.7%)
Systemic anti-infectives
No	498 (22.9%)	66,181 (62.9%)	< 0.001	66,679 (62.1%)
Yes	1672 (77.1%)	39,068 (37.1%)		40,740 (37.9%)

Data pre-processing

Data were corrected for outliers and examined for confounding using correlation analysis and Principal Component Analysis (PCA). Missing values were imputed using multiple imputation by chained equations (MICE), and features with > 55% missing data were dropped [14].

Over the original count of 33 features available, a feature selection was carried out to identify and select those features contributing most to our outcome variable. Feature engineering (variable selection) was evaluated using a RandomForest algorithm to identify the predictive value for each feature, with the default set from 500 to 2500 to assure including most feature combinations [15]. Based on two criteria, being (1) feature had a good predictive value ( > = 10% importance), and (2) feature was expected to have clinical importance, 31 features were included. These consisted of patient characteristics, laboratory values, medication during admission, health care logistics, and comorbidity. Also, more commonly available features (i.e. not specific to PVCP) were included. This was in regard to availability of data and to determine if the model could perform well enough on generic features. More detailed information about the features can be found in the supplementary materials under the ‘Features’ section and in Fig. 1.

Fig. 1.

Results of feature engineering using a RandomForest model. Legend: Feature importance

Sampling methods

The original dataset had an outcome balance where 98% had no PVCP (the major group), compared to the remaining 2% who did (the minor group). The training set was resampled with SMOTE in order to balance the outcome [16]. Resampling was only performed on the training set to prevent information leakage to the test set. The ratio of the outcome class was expressed in percentages of total observations, showing the major group first followed by the minor group. Based on using a ‘ceiling of performance’ metric stating that an outcome should improve more than 10% per resample, the optimum point of resampling was reached at the balance of 70% (major group) and 30% (minor group) and was used as final dataset.

Machine learning algorithms

A RandomForest (RF) classifier was chosen as algorithm to train and predict with, being one of the more well known and commonly used classification algorithms. This invokes better understanding of the predictive model, is easy to configure and implement, and robust in preventing overfitting. Data were split into a train and a test set (70:30 respectively), and the RF model was trained with 10-fold cross validation and tested for performance on the separate test set. Based on data from the department of Cardiology, a validation set was used to evaluate the hospital trained model. Data was extracted from over more than a year after clinical implementation of the model at the department of Cardiology, in order to evaluate clinical utility.

Model performance evaluation

Model performance was evaluated according to the ABCD-framework for evaluation of a clinical prediction model, which includes discrimination with the Area Under the receiver operating Curve (AUC), precision (Positive Predictive Value), recall (True Positive Rate), calibration with a calibration curve, and the overall prediction error with the Brier score [17].

Prospective validation

Based on real-time data, a report on patient level and a dashboard for physicians and staff responsible for quality control was developed to monitor patients and outcomes of the implementation and provided insight per patient, per sub-department, or other needed levels. To ensure clinical utility, implementation of the model was done in EPIC (EPIC, Wisconsin, USA), and output of the model was integrated in the workflow of physicians and nurses. Clinical instructions were provided to the nurses to teach them about the importance of preventing a PVCP and how the model could be applied to daily monitoring. A whitepaper was also used to inform about preventing PVCP using the model and was included in training programs for staff in the department of Cardiology. Please see the section ‘Whitepaper used for internal communication’ in the supplementary materials for more information. Every 3 h the most recent documented patient data were sent to the model, which calculated the risk scores based on the features values. Based on a SQL-query, predictions of the model and events of PVCP in patients were compared in a manual adjudication process to ensure the model’s accuracy. To simplify these interpretations for clinical use, risk scores were shown as a probability score ranging from 0 to 100% using a traffic light colouring in adherence to the risk zones with green for low (0–40% risk), orange for moderate (40–80%, and red for high risk (80–100%) of developing a PVCP in the next 3–24 h. This was also shown in different tabs in the patient file in EPIC, to ensure it was not overlooked. Showing a Best Practice Advisory (BPA) pop-up, personnel were alerted and instructed to prioritize action upon red zone patients. The protocol to be attentive to patients with moderate risk was applied since prevention was the primary goal. When presented with moderate or high risk (orange or red light), the intravenous (IV) cannula was evaluated and if needed removed and reinserted in another point. If the IV cannula had to stay in place, extra daily checks of the IV band-aid and the insertion point were executed by the healthcare personnel. Preferably, the IV cannula was removed and medication was continued using oral medication.

Software

Data pre-processing and sampling were performed using R Version 4.0.2, and R-studio Version 1.3.1073 (R-Studio, Boston, MA, USA). Modelling and evaluating the model performance was performed using Python version 3.9.8. with the scikit-learn version 1.1.1. All Python and R code will be made available immediately following publication to anyone who wishes to access the code, and access requests can be made to the corresponding author.

Results

The developed model was trained and tested using generic hospital data, and a validation set for evaluation of generatability based on the data from the department of Cardiology. The performance of AUC was 0.89 (CI: 0.87–0.91) for the test set, and 0.72 (CI: 0.66–0.78) for the validation set. Precision for the test set and validation set both were 0.94 (CI: 0.94–0.95), indicating a good generalizability to timely predict PVCP. This also applied to recall, with scores of 0.96 (CI: 0.95–0.96) for the test, and validation sets. Brier scores for the test set were 0.05 (CI: 0.04–0.05), and 0.11 (CI: 0.09–0.13) for the validation set. Due to resampling the data, a distorted effect can be observed in the calibration curves. These show an overestimation of predicting positives as opposed to predicting negatives as seen in Fig. 2. In this scenario there would be an abundance of false positives, patients with a high-risk prediction while a low risk is observed. To overcome the overestimation of the model in predicting positives, thus creating an overflow of ‘false alerts’ of high-risk scores for patients, a confusion matrix was applied over predetermined risk zones (low, moderate, and high). Using the validation data of Cardiology, the low and moderate zones showed a mean NPV of 88%, with the model correctly predicting a true negative in 88 out of 100 patients with a risk score below 80%. Patients with a risk score equal to or above 80% showed a PPV of 30%, indicating that almost a third of the predictions correctly predicted a true positive in patients who developed PVCP.

Fig. 2.

Calibration curve for model predictions. Legend: Calibration curve with the model’s predicted probabilities on x-axis against the actual outcome (y-axis)

Before the implementation of the model, cases of PVCP were steadily increasing showing a linear trend upwards for the department of Cardiology. After implementation of the model in the department of Cardiology, PVCP cases were monitored, and a downward trend of cases was observed. This was especially noticeable in the long stay admissions, whereas the acute admissions showed no real change over time. In the first period after go-live, a slight delay in effect on outcome data was observed. This lag was most likely caused by communicating changes across a large group of personnel and changing culture and behaviour concerning prevention of a PVCP based on the output of the model. More information can be found in Fig. 3 for all admissions, and in Fig. 4 specified on type of admission.

Fig. 3.

PVCP cases over time. Legend: Overall trend of PVCP cases, with a purple bar indicating the model go-live in May 2021

Fig. 4.

PVCP cases over time per admission type. Legend: Overall trend of PVCP grouped per type of admission, purple bar indicating go-live in May 2021

To compare the results of the implemented model, two time periods were compared, the second half of the year 2021 - due to go-live in the first half year of 2021 – and the second half of the year 2022. The mean per month of the third and fourth quarter of the year 2021 and 2022 of PVCP cases were taken, and statistically tested for significance with an unpaired two-samples t-test. The incidence of PVCP between the last half year of the year 2021 and 2022 showed a statistically significant decline (p-value of 0.01). Grouped per type of admission for the same time period, in the acute admissions showed no statistically significant difference (p-value of 0.68), whereas the long stay admission showed a statistically significant difference (p-value of 0.01). These findings demonstrate a marked reduction in PVCP cases in the Cardiology Department since the model’s implementation, underlining the importance of continued monitoring with a focus on long-stay admissions.

Discussion

Our results show a strong performing model can be developed with hospital data to predict PVCP in the next 3–24 h during clinical admission and shows sufficient generalizability in performance when validated with data from the department of Cardiology. Our findings are in line with previous studies that have developed machine learning algorithms to predict PVCP for early detection. These studies, along with our own, demonstrate the potential of machine learning in health system associated infection prevention and control efforts [9–12]. Our findings of real-world evaluation are promising, as we observed a decline in PVCP in the case department of Cardiology. This decline was statistically significant in long-stay admissions, as opposed to acute admissions where no statistically significant decline was observed. This was in line with our expectations since length of stay in acute admissions is generally too short to develop PVCP, and because length of stay was the most important predictor of a PVCP.

Despite these promising findings, the study has both strengths and limitations. The first strength was the extraction of outcome data (PVCP) using a SQL rule-based query searching through clinical note data instead of extracting discrete datapoints from the patient record file. Since there simply was no discrete documentation in the patient chart stating PVCP occurred during a clinical admission, relying on clinical notes was the only way to extract PVCP cases.

The second strength was the strong performance of the machine learning model developed on general hospital data, and generalizability of that model for a specific department such as Cardiology. This demonstrated that PVCP can be addressed by a machine learning model developed on generic hospital data and applied to specific departments. The third strength was using an elaborate timespan to evaluate clinical utility with in-depth analysis of early identification brought valuable insights. This underlined the importance of implementing such a model in the clinical workflow by creating easy to follow guidelines when there was a need to act. The fourth and last strength, our ‘risk zone’ approach not only addressed calibration issues, but also enhanced the model’s clinical utility. The risk zones created a more focussed approach for nursing staff, instead of only creating more routine checkups. An example of a more focussed approach was inspecting under the IV-patch for swelling or redness when a patient showed a high-risk score. When in doubt, nursing staff contacted the supervising physician earlier to determine actions to prevent PVCP.

We acknowledge two important limitations of this study, with the first being that evaluation of clinical utility was conducted solely in the Cardiology department. Although findings from other departments are promising, our specific evaluation context may limit the model’s generalizability to other specialties and settings—an observation consistent with prior research [18, 19]. Future research should explore the performance of such a model in diverse clinical contexts to ensure its applicability across various departments. Specifically concerning calibration of the model in light of generalizability, other departments also should analyse high risk predictions and apply risk zones to prevent overestimation of positives. The second limitation was that while the descending trend in PVCP suggests the model’s effectiveness, it is crucial to consider other factors that may have influenced infection rates during the study period. We observed an increase in attention and awareness directed to PVCP, resulting in more communication between health care professionals considering patient risks and the current need for an IV cannula. The model and especially all automated extraction of patient level information seemed to provide most insights to personnel to provide in early identification. This suggests that a machine learning model with a strong performance is mostly used as a red-flag system, while healthcare professionals are better positioned to reduce PVCP cases when prompted by such a system.

Furthermore, the potential impact on therapy and patient follow-up due to PVCP predictions should be considered. Early identification of patients at high risk for PVCP not only supports timely preventive measures, but may also influence clinical decision-making regarding IV therapy, choice of catheter type, or duration of use. Additionally, predictive insights could inform follow-up protocols, prompting post-removal monitoring or patient education to recognize early signs of infection. Integrating predictions into clinical workflows may amplify the effectiveness of PVCP prevention strategies and improve patient outcomes.

Conclusions

Our study establishes the viability of employing hospital data to construct a robust and generalizable machine learning model for the early detection of PVCP. The early identification of such incidents proves beneficial in reducing the risk of PVCP significantly for patients with long-stay admissions in the Cardiology department. This underscores the potential of machine learning applied in clinical practice as a prediction tool to aid infection prevention and control endeavours in optimizing patient care. We recommend that future studies focus on investigating the attribution of a machine learning model and the reduction of PVCP cases in clinical settings.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

K had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. R, F, EP, MPJ, and JN contributed equally as co-authors.

Funding

This study did not receive financial support.

Data availability

Data that support the findings of this study are not publicly available, due to privacy regulation and restrictions on patient data. Hence, the data are not publicly available. However, all code written and applied for this study is available. For more details, and access, Welvaars can be contacted at [k.welvaars01@umcg.nl].

Declarations

Ethics approval and consent to participate

Data was derived from the OLVG, Amsterdam, the Netherlands. The database is de-identified, and approval was granted by the local institutional review board Authorisatie Commissie Wetenschappelijk Onderzoek (ACWO) under reference number 23.020. A clinical trial number is not applicable for this study. Informed consent from all the participants in the study was not deemed as necessary by the ACWO, due to the large number of observations in the dataset. For that same reason, the need for consent to participate was deemed unnecessary by the ACWO. With this, all methods were carried out in accordance with privacy guidelines and regulations. Furthermore, our study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.