The findings of this study suggest that there are differences in the prevalence of ADRs during hospitalization in older inpatients according to dementia status and ADR definition approach. There were no differences in the medication classes implicated in ADRs according to dementia status and ADR definition methodology. Several possible risk factors of ADRs in hospital were identified, including length of hospitalization, the number of regular PIMs on admission, the regular use of a diuretic on admission, marital status listed as missing, unknown or declined to respond and living status prior to admission listed as assisted living or other.

Among all inpatients, a total of 8.3% of those with dementia had an ADR defined by ICD-10-AM, compared with 14.6% patients with no dementia diagnosis (p < 0.001). Additionally, the odds of having an ADR were lower in people with a dementia diagnosis, compared with people without a dementia diagnosis (OR 0.56, 95% CI 0.38, 0.82). The frequency of previous ADRs, prior to hospital admission was also lower in patients with dementia (33.3%), compared with patients without dementia (44.4%). These findings contrast the widely held view that ADRs are more prevalent in people with dementia [14, 38]. Despite this, limited studies have compared the prevalence of ADRs among inpatients with and without dementia. Hofer-Dueckelmann et al. reported that 15.3% of patients with dementia experienced an ADR at admission, compared to 5.9% of patients without dementia (p < 0.001) [39]. However, Onder et al. reported that cognitive impairment was associated with a reduced risk of ADRs occurring during hospitalization (OR 0.70, 95% CI 0.60, 0.83), affirming the findings of our study [40]. Dementia itself is often underdiagnosed and our findings are suggestive of possible under-reporting in this population, therefore the under-reporting of ADRs in those with dementia is twofold [41, 42]. Physicians may underdiagnose ADRs due to the reduced ability of patients with dementia to recognise, report and recall adverse reactions and collaborate with health professionals [14, 40, 43]. Patients with dementia also tend to under-report common symptoms unrelated to their cognitive impairment [44].

With respect to ADR methods, we found significant differences in ADR prevalence, with ICD-10-AM methodology resulting in under-reporting of ADRs compared with research pharmacist classification. Similarly, a meta-analysis reported the prevalence of ADRs in older inpatients to be only 6.0% (95% CI 3.56, 8.88) when detected by ICD-10 coding, in comparison with 8.8% (95% CI 7.46, 10.28) when detected by researchers and 12.6% (95% CI 8.18, 17.76) when detected by treating physicians [5]. Interestingly, in our study, ICD-10-AM identified all patients who had an ADR classified as definite, according to the Naranjo algorithm. We used a more expansive definition of ADRs, encompassing all Naranjo scores, as understanding the potential prevalence of ADRs is important. Intervention by a health professional, including pharmacists, likely contributes to better detection of ADRs [45].

To our knowledge, no studies have directly compared the medication classes implicated in ADRs in those with and without dementia. In those with dementia, the commonest five medication classes implicated in ADRs defined by ICD-10-AM were analgesics, antithrombotic agents, anti-infectives for systemic use, psycho-analeptics and anti-epileptics. Similarly, one study reported that the medication classes most commonly implicated in ADRs on admission in older inpatients with dementia were antibiotics, anti-hypertensives, non-steroidal anti-inflammatory agents and psychotropics [15]. Comparatively, we found in those without dementia, the medication classes most commonly implicated in ADRs were antithrombotic agents, analgesics, anti-infectives for systemic use, anti-hypertensives and psycholeptics. A meta-analysis which reported ADRs in older people at admission found that analgesics, beta-blockers, antibiotics and oral anticoagulants were commonly implicated [5].

We also aimed to identify possible risk factors of ADRs during hospitalization in older inpatients. The model C-statistic was 0.68, which is comparable with that of other ADR prediction tools when applied to their derivation cohorts, such as Adverse Drug Reaction Risk in Older Persons (ADRROP) tool (0.62) [23], the pProspective study to develop a model to stratify the RIsk of Medication-related harm in hospitalised Elderly patients (PRIME) tool (0.69) [22], the GerontoNet tool (0.71) [20] and the Brighton Adverse Drug Reactions Risk (BADRI) model (0.74) [21]. A strong predictor of ADRs was a longer length of hospital stay (OR 1.01, 95% CI 1.01, 1.02). Similarly, a length of hospital stay > 12 days was associated with the development of an ADR in the BADRI model (OR 2.27, 95% CI 1.35, 3.83) [21]. Length of hospital stay may serve as a proxy for the severity of patients’ diseases, possibly reflecting a greater medication burden and thus a heightened risk of ADRs [39]. A larger number of regular PIMs on admission was also associated with ADRs during hospitalization (OR 1.17, 95% CI 1.00, 1.38), aligning with the ADRROP tool where the use of one PIM (OR 1.47, 95% CI 1.10, 1.97) and the use of two or greater PIMs (OR 2.69, 95% CI 1.98, 3.66) was associated with ADRs [23]. The regular use of a diuretic on admission was also a strong predictor of ADRs (OR 1.70, 95% CI 1.24, 2.31). Whilst this covariate has not been included in an existing predictive ADR risk model, it has been reported that diuretics are commonly implicated in ADRs in older people [46]. Marital status listed as missing, unknown or declined to respond (OR 5.90, 95% CI 1.52, 22.83) and living status prior to admission listed as assisted living or other (OR 1.94, 95% CI 1.19, 3.17) were also statistically significant risk factors of ADRs. A lack of family support has been linked to poor adherence with medication regimes in older people, potentially leading to ADRs [47, 48]. However, a small number of patients were present in marital status and living status categories, therefore this statistical significance should be interpreted with caution.

In addition to dementia diagnosis, as previously discussed, several other covariates were found to reduce the odds of an ADR. One such variable was increasing age (OR 0.96, 95% CI 0.94, 0.99). Age > 70 years was a significant predictor of ADRs in hospital in the ADRROP model (OR 1.56, 95% CI 1.05, 2.32) [23], although this is of limited applicability to our population who were all aged ≥ 75 years. Other studies have reported that age is not a significant predictor of ADRs in hospital [49, 50]. Although medication metabolism is altered with age, the medications and dosages used in older patients are often adjusted to compensate for age [50]. We also found that the odds of having an ADR were significantly reduced by the presence of liver disease (OR 0.23, 95% CI 0.07, 0.74), aligning with the findings in the PRIME study, which may be due to caution taken in patients with liver disease, regarding medication choice and dosage [22]. This is contrary to the GerontoNet tool (OR 1.36, 95% CI 1.06, 1.74) [20] and ADRROP tool (OR 2.26, 95% CI 1.31, 3.90) [23], since declining liver function often reduces the excretion of medications, increasing the risk of side effects [51].

There are several strengths of this study. Firstly, this was a multi-centre study of consecutively admitted patients, offering a representative sample of older patients admitted to general medicine, geriatric medicine and rehabilitation at several Sydney hospitals. Secondly, this study involved the use of several reliable, validated tools to define and characterise ADRs, including WHO-ART [27], ATC coding [28], the TGA criteria [29] and the Naranjo algorithm [30], reducing the subjectivity of clinical judgement. Thirdly, the approach to data extraction was comprehensive and systematic, with a trained research pharmacist extensively searching through the eMR to obtain demographic and clinical information. Additionally, the approach to the inclusion of variables in the binary logistic regression, using the AIC metric [34], was validated and systematic.

However, there are several potential limitations of this study. Firstly, classification and characterisation of ADRs by a research pharmacist was only conducted for a sample of 600 patients from the original TO HOME study cohort of 2000 patients, potentially reducing the validity of results. However, the characteristics of this subset were broadly representative of the total cohort. In addition, TO HOME study data were collected during 2016–2017 time period, as such, future studies should replicate the findings using more recent data and Beers Criteria. Secondly, this study utilised data which was collected retrospectively from medical records by a single investigator, potentially resulting in misclassification bias on the prevalence of dementia for all 2000 patients, and the prevalence and severity of ADRs for the sample of 600. Retrospective data collection may has also reduced validity of causality assessment using the Naranjo algorithm, since several items are rarely recorded in the hospital inpatient medical records and therefore could not be assessed, such as re-testing of the potentially causative medication and a obtaining a history of previous adverse reactions. Thus, the proportion of patients with a definite or probable ADR may be underestimated. In addition, the C-statistic was not a strong predictor of ADRs, which is not surprising as multiple factors may explain the suspected ADRs in this population. Thirdly, the cross-sectional study design limits the causality implications. Finally, the cohort study excluded patients who died during hospital admission and participants were only from hospitals in Sydney, Australia, which may not be generalisable to other populations and settings.

This study highlights the need for further research into the prevalence of ADRs during hospitalization among older adults with dementia. Our findings that those with dementia have a reduced prevalence of ADRs and that dementia is associated with a reduced likelihood of having an ADR, contribute to existing discrepancies in the literature and prompt investigation into the possible under-detection of ADRs in this population. Our study also highlights the exacerbation of under-reporting by using manual coding ICD-10-AM codes that define drug-induced diagnoses have limited scope [52]. Suboptimal reporting of ADRs distorts the perceived clinical and financial burden of ADRs on the healthcare system and hinders evaluation of interventions to reduce ADRs [52]. The clinical, economic and public health implications of ADRs are numerous, therefore the improvement of ADR definition systems should be a healthcare priority [52]. Our study suggests that a shift towards detection and reporting of ADRs by health professionals may result in a more accurate portrayal of their prevalence. Pharmacists are optimally positioned to routinely report ADRs in hospital. However, perceived barriers such as heavy workload, understaffing and lack of support by hospital management are problematic and indicate that cultural changes are required at a system level [53]. Computer-assisted coding removes many of these barriers, hence its common use, despite compromised accuracy. Therefore, refinement of computer-based systems to report ADRs represents an alternative approach. Various algorithms and trigger tools have recently been developed, using the eMR data [54,55,56]. Furthermore, coders who record ICD-10-AM codes only have access to information written in the eMR; therefore, this study indicates that clinicians should be more explicit in documenting potential ADRs [57]. This study also highlights the need for further investigation into the key risk factors of ADRs during hospitalization. Since the increasing prevalence of dementia places a growing demand on hospitals to prevent and better manage ADRs in this population, the use of the prediction tools should also be tailored to be sensitive and specific for patients living with dementia [58].