Ageing

Older adults are at a high risk of suboptimal medication use, which includes overuse, underuse and misuse of medications, and this can lead to adverse drug events (ADE) and adverse health outcomes. The factors that lead to suboptimal medication use by older adults can be attributed to ageing-related factors, such as multimorbidity, frailty and changes in PK and PD. Prescribing medications must be a precise balance between minimising the number of medications and using all medications that will be beneficial at the optimal dose for the patient, whilst accounting for age-related factors (Steinman et al., Reference Steinman, Landefeld, Rosenthal, Berthenthal, Sen and Kaboli2006).

Multimorbidity, defined as the presence of multiple concurrent medical conditions, is more common with age and is associated with high mortality, reduced functional status and increased hospitalisation. In a cross-sectional study conducted in Scotland, 42.2% of all patients had one or more morbidities and 23.2% were multimorbid (Barnett et al., Reference Barnett, Mercer, Norbury, Watt, Wyke and Guthrie2012). By 2035, approximately 17% of the UK population is projected to have four or more chronic conditions (Pearson-Stuttard et al., Reference Pearson-Stuttard, Ezzati and Gregg2019). The challenge with multimorbidity and medication management comes with the application of clinical guidelines; most clinical guidelines are built on evidence-based medicine and are designed for the treatment of single diseases, and often overlook medication management in multimorbid older adults. Application of single-disease clinical guidelines in multimorbid older adults can lead to overtreatment. A Norwegian qualitative study with general practitioners (GPs) that explored experiences with and reflections upon the consequences of applying multiple clinical guidelines in older multimorbid adults, found that when GPs focus on person-centred care and refrain from complying with clinical guidelines the risks associated with polypharmacy and overtreatment can be reduced (Austad et al., Reference Austad, Hetlevik, Mjolstad and Helvik2016).

Like multimorbidity, frailty adds more complexity to precisely balancing medication management for older adults. Frailty is a complex geriatric syndrome and a state of vulnerability, which can result in decreased physiological reserve (Clegg et al., Reference Clegg, Young, Iliffe, Rikkert and Rockwood2013). Frailty is common in later life, with prevalence between 10 and 14% for community-dwelling older adults, and up to 50% for older people living in residential aged care facilities (Collard et al., Reference Collard, Boter, Schoevers and Oude Voshaar2012; Kojima, Reference Kojima2015). A systematic review that analysed the evidence and interplay between polypharmacy and frailty in older adults, identified that the association between frailty and polypharmacy may be complex and bidirectional, but polypharmacy is recognised as a major contributor to the development of frailty (Gutierrez-Valencia et al., Reference Gutierrez-Valencia, Izquierdo, Cesari, Casas-Herrero, Inzitari and Martinez-Velilla2018). Many tools have been developed to help clinicians identify inappropriately prescribed medications in older people with polypharmacy (e.g., STOPPFrail) (Thompson et al., Reference Thompson, Lundby, Graabaek, Nielsen, Ryg, Sondergaard and Pottegard2019). However, limited clinical studies demonstrate improvements in frailty when deprescribing medications (Ibrahim et al., Reference Ibrahim, Cox, Stevenson, Lim, Fraser and Roberts2021). A study conducted in aged mice with chronic polypharmacy found that deprescribing high-risk medications attenuated frailty, identifying that there is potential to reduce the effects of frailty by appropriately and precisely managing polypharmacy (more details provided in section ‘Limitations of evidence from human studies and role of preclinical models’) (Mach et al., Reference Mach, Gemikonakli, Logan, Vander Wyk, Allore, Ekambareshwar, Kane, Howlett, De Cabo, Le Couteur and Hilmer2021a).

The physiological changes that occur in ageing can affect PK and PD, which in turn can impose considerable variability in medication management for older adults with polypharmacy. Age-related changes in PK/PD include changes in drug absorption from the gastrointestinal tract, plasma protein binding, drug distribution, reduced hepatic metabolism and clearance, altered renal function, changes to receptors and voltage-gated channels, and changes to the autonomic nervous system (Hilmer et al., Reference Hilmer, Mclachlan and Le Couteur2007b). These changes may be exaggerated in frail older people, although the data is very limited (Hilmer and Kirkpatrick, Reference Hilmer and Kirkpatrick2021). Changes in PK/PD also may make older adults with polypharmacy more vulnerable to ADEs. Given the wide variability in response to medicines by older adults, with the added complexity of under-representation of older adults in PK and PD studies and limited clinical trial data, there has been recent debate about the role of PK and PD studies for frail older adults to inform medication management (McLachlan et al., Reference McLachlan, Hilmer and Le Couteur2009; Mangoni et al., Reference Mangoni, Jansen and Jackson2013; Liau et al., Reference Liau, Lalic, Sluggett, Cesari, Onder, Vetrano, Morin, Hartikainen, Hamina, Johnell, Tan, Visvanathan and Bell2021). There is also scope for PK/PD monitoring, including therapeutic drug monitoring, in clinical practice to improve precision medicine in highly variable older people with polypharmacy.

Interactions

Precision medicine in polypharmacy needs to consider the increased complexity of drug interactions when multiple drugs are used concurrently, particularly in people with multimorbidity. These include PK and PD DDIs, drug–gene interactions, drug–disease interactions, drug–food interactions, drug–lifestyle interactions and drug–microbiome interactions (Johnell and Kiarin, Reference Johnell and Kiarin2007; Guthrie et al., Reference Guthrie, Makubate, Hernandez-Santiago and Dreischulte2015). The challenges of each of these factors as they relate to polypharmacy are described below. Furthermore, these interactions influence each other, which will require advanced bioinformatics to fully understand and predict the overall effect on an individual patient. This concept is illustrated in the case described in Box 1.

Polypharmacy is the greatest risk factor for DDIs, which may be PK, PD or both. DDIs can occur between prescribed drugs, over the counter drugs, complementary and alternative medicines; and affect drugs used acutely, chronically and intermittently. Any change in drug or dose (including prescribing or deprescribing) can impact existing DDIs. As described in section ‘Use of machine learning for predicting polypharmacy interactions and effects’, drug interactions are generally only assessed for drug pairs, and the effects of interactions beyond drug pairs remain poorly understood. Recent attempts have been made to understand the impact of PD DDIs involving multiple drugs in the setting of polypharmacy, for example through tools to measure anticholinergic burden (Salahudeen et al., Reference Salahudeen, Duffull and Nishtala2015).

Drug–gene interactions are common and there is increasing understanding of the role of pharmacogenomics in determining PK and PD variability. While this is only one of many factors that influence variability in drug response in older adults with polypharmacy, it remains important (Dücker and Brockmöller, Reference Dücker and Brockmöller2019). For example, drug clearance through a particular pathway may be affected by a genetic polymorphism, as well as by other drugs that inhibit or induce the pathway. This is a complex two-way relationship, whereby both factors may work in the same or in opposite directions. Furthermore, in a patient with polypharmacy, there may be multiple polymorphisms affecting multiple pathways, and multiple drugs each using these pathways for clearance. A recent review of the impact of pharmacogenomic testing for PK factors in patients with polypharmacy identified six studies of variable quality, and five reported improved clinical outcomes or reduced drug/health utilisation outcomes (Meaddough et al., Reference Meaddough, Sarasua, Fasolino and Farrell2021).

Drug–disease interactions are extremely common in people with polypharmacy, since it often goes hand in hand with multimorbidity. The concept of ‘therapeutic competition’ addresses the clinical challenge of selecting which condition to treat, when treatment of one of a patient’s conditions worsens another of their conditions. It is estimated that one in five older Americans receive medications that may adversely affect co-existing conditions (Lorgunpai et al., Reference Lorgunpai, Grammas, Lee, Mcavay, Charpentier and Tinetti2014).

Drug–food interactions include a wide range of PK and PD interactions between specific drugs and foods, the effects of overall nutritional state on drug PK and PD, and the effects of drugs on nutrition (Schmidt and Dalhoff, Reference Schmidt and Dalhoff2002). Medications can either stimulate appetite, resulting in obesity, or more commonly in people with polypharmacy, can cause nausea and reduce appetite resulting in malnutrition (Fávaro-Moreira et al., Reference Fávaro-Moreira, Krausch-Hofmann, Matthys, Vereecken, Vanhauwaert, Declercq, Bekkering and Duyck2016). Recent studies have used data mining to predict and evaluate food–drug interactions (Rahman et al., Reference Rahman, Vadrev, Magana-Mora, Levman and Soufan2022). This method is highly applicable to people with polypharmacy.

Drug–lifestyle interactions cover interactions with diverse factors such as alcohol, smoking and exercise. There are well-characterised PK and PD drug interactions with alcohol and smoking, including the impact of therapeutic drugs on the clearance of alcohol and nicotine, and impact of alcohol and nicotine on drug clearance. Therapeutic drugs can increase or decrease exercise capacity through cardiorespiratory effects or neuromuscular effects. Anabolic exercise can be used to counter sarcopaenia induced by drugs such as prednisone, or strength and balance training can be used to reduce susceptibility to falls risk-increasing drugs (The Agency for Clinical Innovation, 2021).

The bidirectional interactions between a wide range of drugs and microbiome have recently been characterised and provide some explanation for previously unexplained inter-individual variability in drug response (Weersma et al., Reference Weersma, Zhernakova and Fu2020). A wide range of therapeutic drugs affects the microbiome in different ways, as do age, sex, disease, frailty, dementia and polypharmacy. Recently the microbiome signature of polypharmacy was characterised in observational studies in older people (Nagata et al., Reference Nagata, Nishijima, Miyoshi-Akiyama, Kojima, Kimura, Aoki, Ohsugi, Ueki, Miki, Iwata, Hayakawa, Ohmagari, Oka, Mizokami, Itoi, Kawai, Uemura and Hattori2022). Interventional studies in mice found changes in microbiome with the single polypharmacy regimen tested, which was partially reversed after deprescribing (withdrawal) in old age (Gemikonakli et al., Reference Gemikonakli, Mach, Zhang, Bullock, Tran, El-Omar and Hilmer2022). More research is needed to understand the additive or synergistic effects of the multiple medications in polypharmacy, along with effects of multimorbidity and other variables on the microbiome.

The interactions between drugs and geriatric syndromes, such as falls, frailty and confusion, are well-recognised in geriatric medicine. Drugs are considered the most reversible causes of these presentations (Avorn and Shrank, Reference Avorn and Shrank2008). Polypharmacy itself is one of the strongest risk factors, with different drug classes more strongly associated with specific geriatric outcomes, often with evidence of a dose response and/or cumulative effects (Hilmer and Gnjidic, Reference Hilmer and Gnjidic2009).

Limitations of evidence from human studies and role of preclinical models

Precision medicine requires consideration of a multitude of factors to tailor medication to an individual. The added complexity of considering the enormous number of combinations of drugs in different polypharmacy regimens, along with different combinations of factors within the individual can be overwhelming in terms of interventional clinical trial design. Application of bioinformatics to this challenge is an emerging strategy, described in section ‘Use of machine learning for predicting polypharmacy interactions and effects’.

Randomised trials have investigated the effects of polypharmacy for single diseases. For example, the use of multiple drugs is endorsed in guidelines for conditions such as tuberculosis, HIV, heart failure, ischaemic heart disease and diabetes. Many of the large randomised controlled studies that inform these guidelines include subgroup analyses by age, sex, comorbidities and more recently by frailty (Dewan et al., Reference Dewan, Jackson, Jhund, Shen, Ferreira, Petrie, Abraham, Desai, Dickstein, Køber, Packer, Rouleau, Solomon, Swedberg, Zile and Mcmurray2020; Nguyen et al., Reference Nguyen, Harris, Woodward, Chalmers, Cooper, Hamet, Harrap, Heller, Macmahon, Mancia, Marre, Poulter, Rogers, Williams, Zoungas, Chow and Lindley2021). Another source of evidence has been subgroup analyses of clinical trials investigating treatment of a single disease according to baseline polypharmacy in the participants (Jaspers Focks et al., Reference Jaspers Focks, Brouwer, Wojdyla, Thomas, Lopes, Washam, Lanas, Xavier, Husted, Wallentin, Alexander, Granger and Verheugt2016). This gives information on the effects of polypharmacy on the efficacy and safety of monotherapy for a single disease, but such analyses have not extended to consider the impact of other factors that inform personalised medicine.

Factors that influence the effects of polypharmacy can be evaluated indirectly through observational studies in populations of older adults. The ability of this data to inform precision medicine is currently very limited, with a recent review highlighting the lack of data even on the effects of sex and gender on polypharmacy outcomes, let alone the myriad of other individual factors (Rochon et al., Reference Rochon, Petrovic, Cherubini, Onder, O’Mahony, Sternberg, Stall and Gurwitz2021).

Interventional trials of polypharmacy that consider different baseline characteristics in subgroups would give data comparable to the data that informs precision medicine for monotherapies. It is not ethical or feasible to conduct interventional randomised controlled trials to evaluate the effects of polypharmacy used for multimorbidity in older adults. Recently, a polypharmacy mouse model was developed, to understand the effects of polypharmacy on key outcomes in old age, and to investigate the effects of common factors that might influence these effects, such as the composition of the polypharmacy regimen, age and sex. An assay to measure pharmacokinetic variability as a factor in personalised medicine was also developed (Mach et al., Reference Mach, Wang and Hilmer2021b). Application of this model by our laboratory and international collaborators, has shown that polypharmacy causes frailty and functional impairment, with greater effects seen with drug regimens with higher anticholinergic and sedative load (measured using Drug Burden Index, DBI), greater impairment in old age, and different patterns of physical and cognitive impairments between males and females. These findings are shown in Table 1. There are now opportunities to analyse proteomics, transcriptomics, metabolomics and microbiome data from these well-characterised phenotypes, using systems biology, to identify biomarkers that predict PK and PD responses to polypharmacy and deprescribing. These could be used to inform precision medicine for people with polypharmacy, for example, by integration into physiological-based PK-PD modelling.

Table 1. The effects of polypharmacy on global health outcomes in a mouse model: impact of drug regimen, age and sex

Note: Drugs administered in polypharmacy regimens were *simvastatin, metoprolol, omeprazole, paracetamol, citalopram; **zero DBI polypharmacy: simvastatin, metoprolol, omeprazole, paracetamol, irbesartan; low DBI polypharmacy: simvastatin, metoprolol, omeprazole, paracetamol, citalopram; high DBI polypharmacy: simvastatin, metoprolol, oxybutynin, oxycodone, citalopram; each drug from high DBI regimen also administered as monotherapy; ^ simvastatin, metoprolol, aspirin, paracetamol, citalopram.

Use of decision support tools

Criteria included in existing decision support tools and limitations

Several decision support tools and guidelines have been developed to optimise polypharmacy and reduce medication-related harm in older adults. Some tools simply provide a list of potentially inappropriate medications (PIM) in the older population such as the PRISCUS list (Latin for ‘old and vulnerable’) (Holt et al., Reference Holt, Schmiedl and Thürmann2010). In contrast, other tools have additional criteria for identifying PIMs, such as interaction between drugs and diseases for example the Beers criteria (Fick et al., Reference Fick, Semla, Steinman, Beizer, Brandt, Dombrowski, Dubeau, Pezzullo, Epplin, Flanagan, Morden, Hanlon, Hollmann, Laird, Linnebur and Sandhu2019) and Screening Tool of Older Persons’ Prescriptions (STOPP) and Screening Tool to Alert to Right Treatment (START) (O’Mahony et al., Reference O’Mahony, O’Sullivan, Byrne, O’Connor, Ryan and Gallagher2015). It is important to consider patient-specific factors such as dose appropriateness for the particular patient. However, a systematic review of different polypharmacy tools (Masnoon et al., Reference Masnoon, Shakib, Kalisch-Ellett and Caughey2018) found that whilst 64.3% of tools mention dosing, only 2.4% consider specific doses being used, such as the DBI (Hilmer et al., Reference Hilmer, Mager, Simonsick, Cao, Ling, Windham, Harris, Hanlon, Rubin, Shorr, Bauer and Abernethy2007a). Development of electronic Clinical Decision Support Systems (CDSS) has been identified as a key facilitator in uptake of these tools in busy clinical practice and a step towards precision medicine in polypharmacy.

Table 2 summarises different electronic CDSS published in the last 10 years, based on existing polypharmacy tools. Studies were identified (Mouazer et al., Reference Mouazer, Tsopra, Sedki, Letord and Lamy2022) and data were extracted (Masnoon et al., Reference Masnoon, Shakib, Kalisch-Ellett and Caughey2018) using the search strategy utilised in previous literature, with the date range set to the last 10 years (January 2012 to October 2022).

Table 2. Electronic clinical decision support systems to optimise polypharmacy

Note: Studies were identified using the search strategy outlined by Mouazer et al. (Reference Mouazer, Tsopra, Sedki, Letord and Lamy2022), with the date range was set to the last 10 years (2012 to October 2022). Data items included in the table were guided by Masnoon et al. (Reference Masnoon, Shakib, Kalisch-Ellett and Caughey2018) and Mouazer et al. (Reference Mouazer, Tsopra, Sedki, Letord and Lamy2022). Y, Yes (characteristic considered by the CDSS); N, No (characteristic not considered by the CDSS), Y^a, mentions dosing only; Y^b, based predominantly on actual doses being used.

Abbreviation: ACB Scale, Anticholinergic Cognitive Burden Scale; DBI; Drug Burden Index; DDI; Drug–drug interaction; DDSI, Drug–disease interaction; DGI, Drug–gene interaction; MRCI, Medication Regimen Complexity Index; NS, not specified (unclear if characteristic considered by the CDSS); OPERAM, Optimising Therapy to Prevent Avoidable Hospital Admissions in Multimorbid Older Adults; PIM, Potentially Inappropriate Medication; PRISCUS, Latin for ‘old and vulnerable’; rPATD, Revised Patients’ Attitudes Towards Deprescribing; START, Screening Tool to Alert to Right Treatment; STOPP, Screening Tool of Older Persons’ Prescriptions.

In terms of criteria considered by different electronic CDSSs to guide polypharmacy review, all tools require a patient’s medication list (Holt et al., Reference Holt, Schmiedl and Thürmann2010). Some tools apply other additional criteria to tailor the output to the specific patient, such as health conditions or disorders, laboratory test results and pharmacogenomic data (Mouazer et al., Reference Mouazer, Tsopra, Sedki, Letord and Lamy2022). For example, the Software ENgine for the Assessment and optimisation of drug and non-drug Therapy in Older peRsons (SENATOR) uses STOPP START (O’Mahony et al., Reference O’Mahony, O’Sullivan, Byrne, O’Connor, Ryan and Gallagher2015), the MedSafer system uses Beers criteria (Fick et al., Reference Fick, Semla, Beizer, Brandt, Dombrowski, Dubeau, Eisenberg, Epplin, Flanagan, Giovannetti, Hanlon, Hollmann, Laird, Linnebur, Sandhu and Steinman2015), STOPP(O’Mahony et al., Reference O’Mahony, O’Sullivan, Byrne, O’Connor, Ryan and Gallagher2015) and evidence-based recommendations from Choosing Wisely Canada (McDonald et al., Reference McDonald, Wu, Rashidi, Forster, Huang, Pilote, Papillon-Ferland, Bonnici, Tamblyn, Whitty, Porter, Battu, Downar and Lee2019; Baysari et al., Reference Baysari, Duong, Hooper, Stockey-Bridge, Awad, Zheng and Hilmer2021), and the Goal-directed Medication review Electronic Decision Support System (G-MEDSS) uses The DBI Calculator (Kouladjian et al., Reference Kouladjian, Gnjidic, Chen and Hilmer2016; Kouladjian O’Donnell et al., Reference Kouladjian O’donnell, Reeve and Hilmer2022).

There are limitations in terms of criteria considered by existing tools. Firstly, real-world patients are complex, with multiple conditions and medications. Current tools however lack intelligent algorithms to account for multiple complex interactions in a patient, for example, different drug–food and drug–gene interactions (Finkelstein et al., Reference Finkelstein, Friedman, Hripcsak and Cabrera2016; Mehta et al., Reference Mehta, Kochar, Kennelty, Ernst and Chan2021; Westerbeek et al., Reference Westerbeek, Ploegmakers, De Bruijn, Linn, Van Weert, Daams, Van Der Velde, Van Weert, Abu-Hanna and Medlock2021; Damoiseaux-Volman et al., Reference Damoiseaux-Volman, Medlock, Van Der Meulen, De Boer, Romijn, Van Der Velde and Abu-Hanna2022; Mouazer et al., Reference Mouazer, Tsopra, Sedki, Letord and Lamy2022). Additionally, an important consideration in precision medicine is distinguishing between theoretical and clinically relevant drug interactions for a particular patient, which is another limitation. Caring for real-world patients often requires managing conflicting recommendations from different guidelines in the same patient but existing tools lack algorithms to provide tailored decision support in these scenarios. Whilst there is data suggesting that machine learning, which refers to the use of algorithms and statistical models to analyse and interpret medical data in order to generate predictions or insights that can inform clinical decision-making, may be a promising approach to developing polypharmacy CDSS (Corny et al., Reference Corny, Rajkumar, Martin, Dode, Lajonchère, Billuart, Bézie and Buronfosse2020), previous research has stated that most current electronic CDSS have not used machine learning algorithms to target output signals (Mouazer et al., Reference Mouazer, Tsopra, Sedki, Letord and Lamy2022). Lastly, an important aspect of precision medicine is making therapeutic decisions whilst considering the patient’s specific goals of care. However, goals of care are not routinely considered by different polypharmacy CDSS (Finkelstein et al., Reference Finkelstein, Friedman, Hripcsak and Cabrera2016; Mehta et al., Reference Mehta, Kochar, Kennelty, Ernst and Chan2021; Mouazer et al., Reference Mouazer, Tsopra, Sedki, Letord and Lamy2022). Recently, some CDSS have integrated goals of care assessment with other tools to address polypharmacy (Mangin et al., Reference Mangin, Lamarche, Agarwal, Banh, Dore Brown, Cassels, Colwill, Dolovich, Farrell, Garrison, Gillett, Griffith, Holbrook, Jurcic-Vrataric, Mccormack, O’reilly, Raina, Richardson, Risdon, Savelli, Sherifali, Siu, Tarride, Trimble, Ali, Freeman, Langevin, Parascandalo, Templeton, Dragos, Borhan and Thabane2021; Kouladjian O’Donnell et al., Reference Kouladjian O’donnell, Reeve and Hilmer2022).

Use of machine learning for predicting polypharmacy interactions and effects

Understanding polypharmacy effects is an essential step to optimise medication regimens. However, most of the known polypharmacy effects are highly variable and non-specific and usually not detectable in clinical trials (Bansal et al., Reference Bansal, Yang, Karan, Menden, Costello, Tang, Xiao, Li, Allen, Zhong, Chen, Kim, Wang, Heiser, Realubit, Mattioli, Alvarez, Shen, Gallahan, Singer, Saez-Rodriguez, Xie, Stolovitzky and Califano2014). Given the vast number of drug combinations, neither experiments nor clinical trials can investigate the effects of DDIs for all possible combinations due to time and cost. Therefore, computational methods have been developed for predicting unknown DDIs that cause effects that cannot be attributed to single drugs alone (Han et al., Reference Han, Cao, Wang, Xie, Ma, Yu, Wang, Xu, Zhang and Wan2021).

Generally, the methods for predicting DDIs are divided into two categories: (1) prediction of DDIs, and (2) prediction of specific types of DDIs (Han et al., Reference Han, Cao, Wang, Xie, Ma, Yu, Wang, Xu, Zhang and Wan2021). The first category predicts whether a pair of drugs will cause DDI. This category can be further classified into similarity-based and classification-based methods. The idea behind the similarity-based method is that if drug A and drug B cause a DDI, drug C similar to drug A should also interact with drug B. Different types of Drug–drug similarities are used based on molecular structures, side effects, pharmacology and biological elements (e.g., carriers, transporters, enzymes and targets) (Gottlieb et al., Reference Gottlieb, Stein, Oron, Ruppin and Sharan2012; Vilar et al., Reference Vilar, Harpaz, Uriarte, Santana, Rabadan and Friedman2012; Ferdousi et al., Reference Ferdousi, Safdari and Omidi2017). In contrast, classification-based methods treat the prediction of DDI between paired drugs as a binary classification task. Known drug pairs with DDI and drug pairs with non-DDI are used as positive and negative cases, respectively, to build classification models such as logistic regression, naïve Bayes, k-Nearest neighbours, and support vector machine (Cheng and Zhao, Reference Cheng and Zhao2014; Huang et al., Reference Huang, Zhang, Qu, Sanseau and Yang2014; Li et al., Reference Li, Huang, Fu, Wang, Wu, Ru, Zheng, Guo, Chen, Zhou, Zhang, Li, Chen, Lu and Wang2015; Kastrin et al., Reference Kastrin, Ferk and Leskošek2018). The second category predicts whether specific DDI will be caused by a pair of drugs. Zitnik et al. (Reference Zitnik, Agrawal and Leskovec2018) proposed a graph convolutional neural network for multi-relational link prediction called Decagon, which is one of the most well-established models. This multimodal graph includes 645 drugs and 19,085 proteins as nodes, and 4,651,131 DDIs, 715,612 protein–protein interactions, and 18,596 drug–protein interactions as edges. The model predicts associations between pairs of drugs and the specific side effects in the pair as a link prediction task. Since the Decagon model was proposed, other models have been developed for specific DDIs prediction (Nováček and Mohamed, Reference Nováček and Mohamed2020; Bang et al., Reference Bang, Ho Jhee and Shin2021; Masumshah et al., Reference Masumshah, Aghdam and Eslahchi2021).

To achieve better prediction accuracy of the models, various information needs to be extracted from multiple sources. Previous studies have reported that information on the presence and severity of DDIs often vary among databases, which affect the result of model performance (Abarca et al., Reference Abarca, Malone, Armstrong, Grizzle, Hansten, Van Bergen and Lipton2004; Wang et al., Reference Wang, Wong, Lightwood and Cheng2010; Saverno et al., Reference Saverno, Hines, Warholak, Grizzle, Babits, Clark, Taylor and Malone2011). For example, the total number of reported DDIs for 12 commonly prescribed drugs was 1,226 in Kyoto Encyclopedia of Genes and Genomes and 1,533 in DrugBank (Ferdousi et al., Reference Ferdousi, Safdari and Omidi2017). Considering that the number of reported DDIs in these databases increases with each update even between previously existing drug pairs, it is not possible to tell whether drug pairs not reported as having DDIs are true negatives or not-yet-known positives (i.e., false negatives). Therefore, special attention needs to be paid to which data sources were used to build and compare prediction models.

There are a few limitations in the current DDI prediction models. First, the current DDI prediction models only consider effects for two drugs. There is little knowledge of interaction effects unique to three or more drugs that do not occur with two or fewer drugs. Given that most patients with multimorbidity are prescribed more than two drugs, DDI prediction for more than two drugs is important. Secondly, the current DDI prediction models do not consider individual-level predictors (e.g., demographic, clinical and genetic information), as well as detailed drug regimens (e.g., administration route and dosage). Considering that the management of complex drug interactions (e.g., DDIs, drug–gene interactions) as combined parameters affecting drug response is a complex task particularly in older patients with polypharmacy, a comprehensive medication review process will be necessary using a multifaceted intervention bundle with accompanying stewardship program.