Critical care areas require frequent administration of high-alert, critical, short–half-life infusions, intravenous nutrition, sedation and analgesia, as well as other infusions that require rigorous maintenance for continuous delivery. Outside of the intensive care unit (ICU), approximately 60% of all patients will receive an intravenous infusion during their stay [1].

Abrupt and unexpected infusion failure may have severe consequences for patients if the medications are critical, with short–half-life infusions [2]. Continued interruptions to infusions and infusions not running “to time” can also lead to subtherapeutic management. For example, patients receiving antibiotics who require therapeutic drug monitoring based on metrics like area under the concentration-time curve and trough levels often need blood draws before and after administration. The documented time of administration and subsequent blood draws are commonly based on the prescribed regimen and not on the actual completion of the infusion [3,4]. Infusion downstream and upstream occlusion alarms, when substantial, may also contribute to alarm fatigue among clinical staff [5,6]. One extensive study showed that venous access occlusion alarms are responsible for 55% of all intravenous infusion pump alarms in neonatal ICUs [2].

The issue of infusion failure and its determinants has not been comprehensively studied in the literature. Existing in vivo studies have focused on mechanical causes at the vascular access device site [7] and incompatibility issues, either between medications [8] or between medications and administration line materials [9].

A vascular access device (VAD) is defined by the Infusion Nurses Society of the United States as a “catheter, tube, or device inserted into the vascular system, including veins, arteries, and bone marrow” [10]. Definitions for VAD failure include situations where the catheter stops working safely before its intended dwell time or before the traditional 72- to 96-hour dwell time limit [11,12]. Recent guidelines from the Centers for Disease Control and Prevention state that peripheral VADs do not need to be electively resited “more frequently than every 72 to 96 hours” [13]. Using these definitions, the VAD failure rate has been suggested to be as high as 63%, with mean and median values of 46% and 43%, respectively, across studies [14].

The VAD failure rate has a fundamental relationship with the administration method, with gravity administration having a VAD failure rate twice that of even simple rate control infusion devices [15,16]. Modern infusion devices with increased accuracy for the detection of downstream occlusion issues would be expected to reduce the VAD failure rate further. The management of vascular access and infusions also results in a substantial nursing workload. The Therapeutic Intervention Scoring System-28 allocates 3 points to “multiple intravenous medications” (ie, more than 1 medication, “either as single shots or continuously”), 3 points to any “single vasoactive medication,” 4 points in the case of “multiple vasoactive medications, regardless of types and doses,” and 2 points for the care of a “central venous line.” Therefore, continuous infusions of critical, short–half-life intravenous medications via a central VAD could consume 5 to 9 points from a maximum workload of 46 points that can be undertaken by 1 nurse [17], equating to 10%-19% of a critical care nurse’s total activity time. In a study on nursing workload in ICUs with an average length of stay of 7.7 days, it was found that the mean score based on the Therapeutic Intervention Scoring System-28 was 23 (range 14-32 points) and that nursing time constituted the largest economic cost for ICUs [18].

A 2019 study [19] indicated an “excessive nursing workload” across ICUs that was significantly associated with quality of care. Reducing the number of interventions nurses need to undertake to avoid infusion interruption and to increase infusion longevity would be expected to reduce the baseline of nursing workload in intensive care, high-dependency units, and lower-acuity care areas.

In a 2021 study [20] on the impact of infusion alerts and alarms on nursing workflow, alarms and alerts from both intermittent and continuous infusions were analyzed. Alerts, such as those generated by the Dose Error Reduction System due to dose or rate selection by the clinician outside of the defined limits for individual medications, do not interrupt infusions. The study deemed alerts and alarms as “undesirable error states” and described specific conditions that would interrupt infusions, such as flow occlusion and air-in-line alarms.

In our study, we developed working definitions for infusion longevity, and conversely, infusion failure as follows: infusion longevity may be described as the length of time during which a continuous infusion runs without an alarm state causing unexpected and unplanned interruption to the infusion and comes to an end as a planned cessation. Infusion failure may be described as an infusion that does not reach a planned cessation without clinician interventions to manage unexpected and unplanned interruptions.

This study aimed to identify and rank determinants of the longevity of continuous infusions delivered by syringe drivers through the use of nonlinear predictive models to evaluate key factors, and subsequently, develop and test a model for predicting the likelihood of successful infusion longevity; this also involves determining the best predictive model for future use. We expected the analysis to show therapeutic practices and pump management processes that may assist with infusion longevity. Additionally, we aimed to determine which medications are more likely to cause infusion failure and may warrant more intense observation or access management. We also sought to identify critical care units in which infusion failure is more likely to occur and to assess the likelihood of uninterrupted infusion that can be expected in these care areas.

We collected infusion data from smart syringe pumps of type CareFusion/BD Alaris Plus CC from different hospitals in Spain. These data are part of a larger data aggregation for the European region, which is held as a repository as part of the obligation for medical device manufacturers to maintain vigilant postmarket surveillance programs for regulatory and quality purposes.

These data are collected passively as part of the standard function of infusion pumps, capturing all events including alerts, alarms, fault conditions, and programming of all infusions. No patient data are recorded. As these data are retrospective, and therefore, cannot influence clinician decision-making, do not record any direct information related to individual patient therapy, and are detached from any patient or clinician information, there was no requirement for formal ethics approval. The Medication Management Solutions Medical Affairs Department of the pumps’ manufacturer gave clearance for using these data in this study. The question of any conflict of interest was also addressed at this stage. None was found, as the variables studied are universal to “smart” infusion pumps and are not exclusive to the pumps studied.

The data set was obtained from 384 pumps and contained information about various variables. These variables include the profile, indicating the hospital care unit or ward; medication name or type; infusion rate; medication concentration; syringe brand and syringe size; occlusion setting, indicating the pressure threshold at which the pump alarms for an occlusion; and a configured category label for a dependent variable, indicating if the infusion ended by an unexpected and unplanned occlusion or as a planned cessation. Table 1 shows the values for each categorical variable in our data set.

We used 5 nonlinear models to fit the data and evaluated them with test data to find the best-fit model. These nonlinear models were the following:

Choosing the best machine learning model to be used in a study among the hundreds of different available models should be based on their characteristics and their previous success in the field. We chose 2 different ensemble models with extreme gradient boosting along with the random forest model. These models differ in use, and they allowed us to combine multiple models to reach a result. These are well-known models that can be used as delegates of ensemble learning methods. We used KNN as a delegate for nonparametric instance-based learning models. SVM was used as the most commonly used kernel-based learning model. Naive Bayes was tested to check a Bayesian learning model with an independence assumption between the predictors. This set of models covered a large area of different learning natures, and the ideal model selection was made based on finding a global optimum. Undertaking a trial-and-error procedure among hundreds of models with infinite parameter selection was beyond the scope of our resources, but selecting a starting set of different models that represented different learning algorithms gave us a diverse and comprehensive starting point.

SVM’s kernel is a radial basis function with the regularization parameter set to 1. XGBoost uses 100 estimators with both the learning rate and maximum depth set to 1. Our random forest uses 100 decision tree estimators, and it uses the Gini index function as its criterion to measure the split quality in each tree. The nearest neighbor (K) was set to 3 for the KNN model. The Naive Bayes model used a Gaussian function. The parameters were tested on a validation set of 20% of the entire data set before running the final output of sample testing.

We then evaluated the performance of each model using an F₁-score and a 5-fold cross validation. F₁-score is a widely used performance metric in classification tasks that measures the balance between precision and recall. It is the harmonic mean of precision and recall, which means that it takes into account both false positives and false negatives, giving equal weight to both [27]. The F₁-score ranges from 0 to 1, where a score of 1 represents perfect precision and recall, and a score of 0 represents poor performance.

The formula for the F₁-score is as follows:

Data were only collected from hospitals in Spain, though the investigational method could be applied to other regions as the software deployed in the smart pumps is available worldwide and the structure and deployment of the medication library from which the data were gathered has been found to be very similar in a previous wide-ranging study of event logs [2].

The study was limited to one type of syringe pump, the CareFusion/BD Alaris CC pump. Similarly, the investigational model can be applied to other pumps, as features like profile, syringe types and brands, medication libraries, and occlusion alarm pressure settings are considered to be universal across syringe pumps. However, one caveat is that the devices in this study have a fairly unique feature, that of in-line pressure monitoring, where the vein pressure of the patient at the point of the VAD is transmitted directly to a dedicated pressure sensor situated downstream of the syringe and in direct line connection to the VAD. The occlusion alarm setting in this device can be set as low as 15 mm Hg above the detected vein pressure, though commonly, this is set at 30 mm Hg above the vein pressure in neonatal care, and some units use an “auto-offset” feature for automatically setting the alarm after 15 minutes of infusion [32]. Generally, syringe driver infusion pumps without in-line pressure monitoring transmit these data from the drive head behind the syringe, which involves added variables of medication viscosity, syringe friction, and a minimum level of pressure detection, which is rarely below 85 mm Hg.

The question of the type of vascular access of each patient in the study was also a limitation, and whether the VAD was central, peripheral, or umbilical in the case of neonates or a “long line,” such as peripherally inserted central catheters, could not be answered given the data available. However, some inferences could be made from clinically acceptable maximum concentrations for peripheral administration from facility protocols. That being said, evidence is accruing that the concentration of administered individual medications (concentration was not identified as a substantial determinant in our study) is not per se as important as the contact time between medications when multiple infusions pass through a single VAD, along with the subsequent interactions between them. [7,8] At low-flow rates, this contact time is extended, with more time for reactions between medications and subsequent precipitate production. In-line filtration to protect the VAD from precipitate occlusion is emerging from the available evidence as an important factor in determining VAD longevity beyond that of VAD type and medication concentration [7,8].

Including more variables to the information directly available from smart pumps in the analysis, such as direct information about the VAD type and other infusions running via one VAD, may provide a more comprehensive understanding of the in vivo factors that influence infusion longevity.

The occlusion alarm setting threshold was the most important variable for infusion longevity, and beyond 2 individual medications, the infusion rate was the next most important variable. The data drill-down (Figure 3) and ratio ranking (Tables 4 and 5) show that lower ratios of occlusion to no occlusion were associated with higher rates of infusion, with the rate bracket of >3.6-≤4.0 mL/hr having the best ratio at 0.311. However, the bulk of infusions in the study run at rates far lower than that, with 60.11% running at below 2.0 mL/hr. This is understandable given the fluid balance (or more correctly, the fluid restriction requirements) of medication infusions in critical care, particularly in neonatology and pediatrics. Due to maintaining critical care patients’ nutrition as well as managing renal failure and fluid balance, it is not uncommon to use higher concentration infusions to deliver continuous infusion doses with smaller volumes.

Considering the findings of this study and in vitro studies of infusion startup delay and infusion “no-flow” interruptions [33], as well as the influence of administration line compliance [34], filters [35], the interplay between multiple infusions [36], and resistance from backcheck and antisyphon valves [35], the risks of protracted and clinically important nondelivery and occlusion are likely at low rates, particularly below rates of 0.5 mL/hr [35]. The study’s findings suggest a balance between the need to restrict fluid delivery to patients and maintaining the integrity and longevity of infusion might be best achieved with a rate ranging from >2.0 to ≤2.4 mL/hr (ratio 0.985), although the next higher rate of >2.4-≤2.8 mL/hr would yield a far better ratio of 0.553, albeit with some compromise in fluid restriction control.

Therefore, we suggest that, when feasible, some relaxation of fluid restriction and medication concentrations be considered to deliver infusions at rates between 2.0 and 2.8 mL/hr and higher, if at all possible, for continuous infusions. The improvement in the occlusion ratio may be related to the simple volume of medication moving through the VAD and “flushing” it more effectively than very low-rate infusions can achieve, or it could be attributed to the previously mentioned concept of reduced contact time between medications being administered through a single VAD at higher rates. It is possible to target this flow rate range even through wider titration ranges by manipulation of the final concentration of medications. The suggested rate range would also assist with the clinical detectability of nondelivery [33,35].

In Figure 2A-I, binary variables linked to the treatment process or unit type are identified. These variables are essentially beyond the direct control or manipulation of clinicians. However, this information remains valuable as a “high-risk” indicator for individual medications that may benefit from concentration manipulation to facilitate higher delivery rates, closer observation of the infusion, or central VAD delivery and exclusive-line administration rather than peripheral administration along with multiple infusions. A multidisciplinary approach to the management of such high-risk medications is advocated.

These findings have important implications for health care professionals who use smart infusion pumps to deliver medications to patients. The study may assist health care professionals to make informed decisions regarding the medication to be administered, concentrations to be used, and infusion duration or rate, to improve infusion longevity, reduce the risk of unplanned infusion interruption, and mitigate risks to the VAD.

The study also highlights the potential of machine learning nonlinear models to predict infusion occlusions in smart infusion pumps. The process of selecting the most appropriate model could be applied to studies involving other medical devices.