Discussion

This study provides novel evidence that ward doctors can identify patients at risk of medication-related harm resulting from PEs. Our findings demonstrate that when clinical ward physician patients selected patients a for a CMR, a significantly higher proportion of these patients were identified with clinically relevant PEs compared to patients in surgical and medical wards where no selection was applied (p<0.001). This effect remained significant after adjusting for confounding in our matched case-control design. These results support the hypothesis that ward doctors can play a crucial role in identifying patients at risk of potential medication-related harm.
Prescribing in the in-hospital setting is associated with PEs (29, 30). The initiation of new medical treatment and adjustment of existing regimens can introduce unanticipated drug-drug interactions, adverse drug events, or unintentional discontinuation of chronic medication (31). Conducting a medication review can help reduce the number of drug-related problems and minimize harm associated with inappropriately prescribed medication. However, medication reviews are labour-intensive and resources for this may be limited, given the projected global shortage of healthcare workers forecast by the World Health Organization (WHO) (32). Selection of high-risk patients would help make the process more efficient and potentially reduce costs. Therefore, it is crucial to use a sensitive and specific prediction tool to allocate these scarce resources to patients in need of PE-mitigating interventions.
The systematic review of Deawjaroen et al. provides a comprehensive overview of 14 currently available prediction tools and their clinical usefulness in detecting adult, hospitalized patients at risk of medication-related harm (18). Interestingly, the authors concluded that none of the tools is optimal, possibly because of heterogeneity in setting, outcome measures, content, and method of development and validation of the prediction tools. Even though this might be a plausible explanation, we would like to propose an alternative hypothesis.
Prescribing medication in the in-hospital setting is a complex process, influenced by multiple human and non-human factors (19, 29). All available tools predicting medication-related harm combine multiple, stand-alone risk factors, such as the number of prescribed drugs (8), ageing or elderly age, impaired renal or hepatic function, and (33) the use of specific high-risk medication, such as methotrexate and non-steroidal anti-inflammatory drugs (NSAIDs) (9, 11-13), amongst others (18). The Medicine Risk Score (MERIS) (17) and Drug-Associated Risk Tool (DART) (33) are examples of prediction tools that use some of these risk factors. Interestingly, all risk factors used in available prediction tools are patient related, suggesting that medication-related harm is predicted by patient-related factors only.
In a previous study, we made an overview of the factors, both protective and facilitating, reported in the literature as influencing in-hospital PEs (29). When these factors were classified in domains, it became clear that domains other than patient-related factors also significantly influenced in-hospital PEs. For example, organization, prescriber, and technology-related factors. Yet none these potential risk factors are included in currently available prediction tools, which might explain the poor performance of these tools.
It therefore seems appropriate to include most, preferably all, factors identified as influencing in-hospital PEs in future prediction tools (29), that is, not only factors facilitating in-hospital PEs, characteristic of a Safety-I approach, but also factors that protect against in-hospital PEs, adopting a Safety-II perspective (34), which better reflects daily practice and thus provides a realistic risk estimation (35). While prediction tools should be easy and user-friendly (28), sensitive, specific, and thus accurate, including all the factors that influence in-hospital PEs might make the tool too cumbersome to use. Artificial Intelligence (AI) is rapidly gaining a role in healthcare delivery and may provide new possibilities in prediction tool development. Nevertheless, developing any prediction tools, including AI-driven, necessitates collecting, interpreting, and statistical analyses of clinical data, which is a complex, labour-intensive, and time-consuming undertaking. Therefore, there might be power in simplicity.
Lynn et al. investigated the role of doctors on wards in identifying patients who might benefit from palliative care services (36). To this end, doctors were asked to identify patients based on reflecting on the ‘“surprise” question (SQ): Would it be surprising for this patient to die in the next year? (or the next few months?)”, with the answer “no’ being a trigger for referral to specialist services. This stratification strategy has since been used in several patient populations (37-39). To our knowledge, little is known about the basis for doctors’ answers this question. Do they consider several factors, such as quantifiable patient-related factors (e.g., frequent admissions (40), as well as non- quantifiable measures? There are perhaps parallels between the reasoning underlying the answer to the surprise question and why ward doctors selected certain patients for a CMR. Factors taken into consideration might include not only patient-related factors, such as a patient’s habit of avoiding care, medication in adherence, questions regarding the ability to keep an overview of medication regimens, ‘over the counter’ medication use, and ordering medication over the internet, but also whether a patient has a social safety network providing care, whether adequate patient information was received when a patient was transferred from another hospital, whether there was medication reconciliation at hospital admission, if the ward was understaffed, or if there were communication difficulties between the healthcare professionals and the patient. Some doctors may have also selected patients based on an ‘off feeling’ about a patient or wanted to check their assessment. Interestingly, doctors appear to take their own abilities into account in their risk prediction. For example, if they felt that their (drug) knowledge, prescribing skills, or experience was insufficient to form an opinion about a patient’s medication regimens, they sought advice (29). Thus to identify patients at risk of medication-related harm, doctors could perhaps ask themselves: “Would I be surprised if there were prescribing errors in this patient’s medication? ”. If the answer is ‘no’, then that patent’s medication list should be reviewed. This simple prediction strategy uses ward doctors as a predictive indicator, and is based on a doctor’s clinical assessment, experience, and expectations. Therefore, this may provide an easier, user-friendly, and accessible solution for the prediction of patients at risk of in-hospital PEs.

Limitations & strengths

This study has both strengths and limitations. This is the first study to suggest that ward doctors can identify patients at risk of PEs. This prediction strategy does not require a priori data collection to identify and statistically combine multiple predictors to estimate the risk of PEs for an individual patient. Compared to currently available prediction tools or AI, this strategy is easier to use in daily practice (28). Unfortunately, due to the observational design of this study, only patients suspected of PEs by the Otolaryngology and Oncology ward doctors were assessed. This means that patients who were not suspected of having PEs were not assessed by the ISP team. A future study should include these patients in order to better determine the sensitivity and specificity of this strategy.