Evaluation of Clinical Usefulness
Figure 2 shows the decision curve, with the range of threshold probabilities for which the AIME model could be clinically useful. The grey lines indicate the net benefit when pharmacist interventions are provided for all patients (“treat all”), or for the horizontal black line, when no patient receives pharmacist intervention (“treat none”). These serve as a reference to judge whether the AIME and the Polypharmacy models are clinically useful (indicated by the positive values above the “treat all” and “treat none” lines). The turquoise line shows the clinical value of the binary Polypharmacy model (with a single predictor) for medication harm. This model shows minimal clinical usefulness when compared with a “treat all” approach. The red lines show the clinical usefulness of the AIME model. Between the threshold probabilities of approximately 0.05 and 0.7 the AIME model has greater net benefit (i.e. clinically useful) than a ‘treat all’ and ‘treat none’ approach, reinforcing that prioritisation of patients with risk above the 5% probability threshold would be a useful approach.
Insert Figure 2

Discussion

Key findings

We report the development and internal validation of the AIME model, for predicting the risk of inpatient medication harm. The model was developed in a quaternary Australian hospital, in the general medical and geriatric setting. The final model consisted of 10 variables, including high-risk medications, laboratory tests, history of medication allergies and hospital utilisation data, extracted from digital databases at the hospital. The AIME model demonstrated reasonable discrimination (AuROC = 0.70), similar to that of previously published models [11-12, 35-36], (AuROC 0.70-0.74), and superior to three existing inpatient models (AuROC 0.59-0.63) [37-39].

The AIME model compared with existing models

Outcome

The selection of a study outcome is crucial to developing a useful model. Currently, there are a limited number of models that predict actual medication harm, with some using medication errors or medication-related problems, as their primary outcome [17]. In this study, only events that were likely to have resulted in clinical or physiological harm to the patient were included. This is important given that around 90% of medication errors cause no patient harm [18]. Therefore, a key focus of prognostic studies should be on outcomes or actual patient harm, versus medication errors where clinical consequences are unknown . In our study, the rate of harm was approximately 7%, which is consistent with local and international data [6, 36, 40]. Causality assessment showed similarities between event likelihood and severity, to the findings of the study by Tangiisuran et al and the BADRI ADR risk score – a European model for inpatient harm [11].
Cardiovascular medications were the most common medication groups leading to inpatient harm (n=36). This was similar to the findings of the GerontoNet risk score where the most common adverse events included cardiovascular and antiarrhythmic complications [41]. Adverse events associated with cardiovascular agents have been highlighted by multiple studies [42-44]. A recent Australian study of 768 hospital admissions identified the use of multiple antihypertensives as the most significant predictor of medication-related hospitalisation [43]. Similarly, a study of inpatient medication harm, using the Institute for Healthcare Improvement’s ADE Trigger Tool, found that cardiovascular agents were the most common drug class association with harm [6]. Given the rise in the prescription of cardiovascular therapies, in particular in the aging population, it is essential that there is a greater focus on these medications [45].
Regular medication review and deprescribing, in particular in older frail adults, where less aggressive blood pressure control is desirable, should be a key focus for clinicians [35]. As half of the events in this study were either possibly or definitely preventable, collaborative medication management, in particular at transitions in care, can play a major role in improving patient outcomes [5]. A model such as the AIME will help identify and prioritise high-risk patients for timely interventions at transition.