Evaluating prediction of short-term tolerability of five type 2 diabetes drug classes using routine clinical features: UK population-based study

Evaluating prediction of short-term tolerability of five type 2 diabetes drug classes using routine clinical features: UK population-based study | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search Evaluating prediction of short-term tolerability of five type 2 diabetes drug classes using routine clinical features: UK population-based study View ORCID Profile Pedro Cardoso , View ORCID Profile Katie G. Young , View ORCID Profile Rhian Hopkins , View ORCID Profile Bilal A. Mateen , View ORCID Profile Ewan R. Pearson , View ORCID Profile Andrew T. Hattersley , View ORCID Profile Trevelyan J. McKinley , View ORCID Profile Beverley M. Shields , View ORCID Profile John M. Dennis the MASTERMIND consortium doi: https://doi.org/10.1101/2025.02.17.25322406 Pedro Cardoso 1 University of Exeter Medical School. Address: Clinical and Biomedical Sciences, RILD Building, Royal Devon & Exeter Hospital , Barrack Road, Exeter EX2 5DW, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Pedro Cardoso For correspondence: p.cardoso{at}exeter.ac.uk Katie G. Young 1 University of Exeter Medical School. Address: Clinical and Biomedical Sciences, RILD Building, Royal Devon & Exeter Hospital , Barrack Road, Exeter EX2 5DW, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Katie G. Young Rhian Hopkins 1 University of Exeter Medical School. Address: Clinical and Biomedical Sciences, RILD Building, Royal Devon & Exeter Hospital , Barrack Road, Exeter EX2 5DW, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Rhian Hopkins Bilal A. Mateen 2 University College London, Institute of Health Informatics. Address: University College London , 222 Euston Rd, London NW1 2DA, London, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Bilal A. Mateen Ewan R. Pearson 3 University of Dundee. Address: Division of Molecular & Clinical Medicine, Ninewells Hospital and Medical School, University of Dundee , Dundee DD1 9SY, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Ewan R. Pearson Andrew T. Hattersley 1 University of Exeter Medical School. Address: Clinical and Biomedical Sciences, RILD Building, Royal Devon & Exeter Hospital , Barrack Road, Exeter EX2 5DW, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Andrew T. Hattersley Trevelyan J. McKinley 1 University of Exeter Medical School. Address: Clinical and Biomedical Sciences, RILD Building, Royal Devon & Exeter Hospital , Barrack Road, Exeter EX2 5DW, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Trevelyan J. McKinley Beverley M. Shields 1 University of Exeter Medical School. Address: Clinical and Biomedical Sciences, RILD Building, Royal Devon & Exeter Hospital , Barrack Road, Exeter EX2 5DW, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Beverley M. Shields John M. Dennis 1 University of Exeter Medical School. Address: Clinical and Biomedical Sciences, RILD Building, Royal Devon & Exeter Hospital , Barrack Road, Exeter EX2 5DW, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for John M. Dennis Abstract Full Text Info/History Metrics Supplementary material Data/Code Preview PDF Abstract Introduction A precision medicine approach in type 2 diabetes (T2D) needs to consider potential treatment risks alongside established benefits for glycaemic and cardiometabolic outcomes. Considering five major T2D drug classes, we aimed to describe variation in short-term discontinuation (a proxy of overall tolerability) by drug and patient routine clinical features and determine whether combining features in a model to predict drug class-specific tolerability has clinical utility. Methods UK routine clinical data (Clinical Practice Research Datalink, 2014–2020) of people with T2D initiating glucagon-like peptide-1 receptor agonists (GLP-1RA); dipeptidyl peptidase-4 inhibitors (DPP4i); sodium-glucose co-transporter-2 inhibitors (SGLT2i); thiazolidinediones (TZD) and sulfonylureas (SU) in primary care were studied. We first described the proportions of short-term (3-month) discontinuation by drug class across subgroups stratified by routine clinical features. We then assessed the performance of combining features to predict discontinuation by drug class using a flexible machine learning algorithm (Bayesian Additive Regression Trees). Results Amongst 182,194 treatment initiations, discontinuation varied modestly by clinical features. Higher discontinuation on SGLT2i and GLP-1RA was seen for older patients and those with longer diabetes duration. For most other features, discontinuation differences were similar by drug class, with higher discontinuation for patients who had previously discontinued metformin, females, and people of South-Asian and Black ethnicities. Lower discontinuation was seen for patients currently taking statins and blood pressure medication. The model combining all sociodemographic and clinical features had a low ability to predict discontinuation (AUC = 0.61). Conclusion A model-based approach to predict drug-specific discontinuation for individual patients with T2D has low clinical utility. Instead of likely tolerability, prescribing decisions in T2D should focus on drug-specific side-effect risks and differences in the glycaemic and cardiometabolic benefits of available medication classes. Key message Routine clinical features are not sufficient to predict individuals likely to discontinue (maybe say tolerate due to the proxy stuff) T2D glucose treatment. Why did we undertake this study? Precision medicine studies aiming to guide the choice of type 2 diabetes treatment have mainly evaluated treatment benefits, meaning there is little evidence to inform targeting based on potential treatment risks. What is the specific question(s) we wanted to answer? Can routine clinical features be used to predict individual patients likely to discontinue therapy short-term (a proxy of overall drug tolerability), considering five major type 2 diabetes glucose-lowering drug classes. What did we find? Routine clinical features are associated with differences in short-term tolerability of the five drug classes, but a model combining features to predict likely tolerability has low predictive utility. What are the implications of our findings? Overall, type 2 diabetes drug tolerability cannot be accurately predicted for individual patients using routine clinical features. Prescribing decisions in type 2 diabetes should focus on drug-specific side-effect risks and differences in glycaemic and cardiometabolic benefits. Introduction The prevalence of diabetes has been steadily increasing worldwide to an estimated 800 million affected adults, with the majority of individuals being affected with type 2 diabetes (T2D), with most of these requiring drug therapies to control glycaemia levels and lower the risk of adverse complications [ 1 , 2 , 3 ]. Managing T2D requires multiple behavioural and pharmacological treatment strategies, including management of glycaemia, weight, cardiovascular risk factors, comorbidities, complications, and medication side-effects [ 4 ]. Metformin is currently prescribed as a first-line medication for treating T2D due to its safety profile, but as T2D progresses, a single therapy becomes insufficient [ 4 ]. T2D treatment guidelines recommend the addition of other glucose-lowering therapies – glucagon-like peptide-1 receptor agonists (GLP-1RA); dipeptidyl peptidase-4 inhibitor (DPP4i); sodium-glucose co-transporter-2 inhibitors (SGLT2i); thiazolidinediones (TZD) and sulfonylureas (SU) – when a single agent becomes insufficient [ 4 ]. Despite recent progress in T2D precision medicine [ 4 ], there is still uncertainty on the best treatment for individual patients and what metrics should be used for treatment targeting. While recent studies have proposed targeting specific T2D medication at the individual-patient level based on differences in glycaemic benefits [ 5 , 6 ], comparable data to inform targeting based on drug side-effects and/or likely tolerability is limited. All glucose-lowering agents have established specific side-effects that can frequently lead to poor tolerability and discontinuation [ 7 , 8 , 9 ]. Whilst the risk of each individual drug-specific side-effects can be studied directly [ 4 , 10 ], an alternative is to evaluate short-term discontinuation risk as a proxy outcome measure of likely tolerability, applicable across all drug classes [ 5 , 11 ]. This approach is supported by recent studies demonstrating that short-term fracture and genital infection side-effects lead to an increased risk of early discontinuation with SGLT2i [ 9 , 12 ]. Although multiple other factors can lead to a perceived lack of medication efficacy or access to medication [ 4 , 13 ], short-term side-effects are likely the main reason for early discontinuation. We aimed to investigate the potential for a precision medicine approach based on routine clinical features to predict which patients are more likely to tolerate each of the five major T2D therapy classes after initial metformin, using 3-month discontinuation as a proxy outcome measure. First, we assessed differences in short-term discontinuation when stratifying by clinical and sociodemographic features and hence highlighting potential treatment heterogeneity. We then evaluated the potential for targeting specific therapy classes in individual patients by combining multiple features in a model to predict drug-specific likely tolerability. Methods Study population Adults with T2D initiating GLP-1RA, DPP4i, SGLT2i, TZD or SU therapies of any type for the first time between 1 January 2014 and 31 October 2020 were identified in the UK population-representative Clinical Practice Research Datalink (CPRD) Aurum dataset [ 14 ] linked to Hospital Episode Statistics (HES), Office for National Statistics (ONS) death registrations and individual-level Index of Multiple Deprivation (IMD), following our previously published cohort profile [ 15 ] (see https://github.com/Exeter-Diabetes/ for all codelists and cohort preparation algorithms). We excluded individuals initiating these therapies as first-line monotherapy (as this does not align with current guidance [ 4 ]), multiple glucose-lowering therapies on the same day (as this is non-standard in the UK), missing baseline HbA 1c and baseline HbA 1c <53 mmol/mol (7%) (as the indication for treatment initiation may not be for glucose-lowering) (sFig.1). Outcome The primary outcome was discontinuation within three months of initiation (a proxy for drug tolerability), with no changes to other glucose-lowering therapies during this period. Therapy discontinuation was defined as a prescription gap above 180 days. To ensure sufficient follow-up to assess this definition, patients with less than six months of follow-up data after their last prescription (due to the end of the study period, practice deregistration, or death) were excluded since discontinuation could not be confirmed. As sensitivity analyses, we also assessed discontinuation when defined at 6- and 12-months post-drug initiation. Candidate predictors Candidate predictors were selected to represent routinely available sociodemographic, behavioural and clinical features [ 5 , 6 ]. These comprised: current age, duration of diabetes, sex (self-reported, categorised into male, female), social deprivation (English Index of Multiple Deprivation quintile), ethnicity (self-reported, categorised into major UK groups: White, South Asian, Black, Mixed, other), smoking status (self-reported, categorised into active smoker, ex-smoker, non-smoker), baseline HbA 1c (closest to treatment start date; range in previous six months to +7 days), BMI, eGFR (closest to treatment start date; range in previous 2 years to +7 days), number of ever prescribed glucose-lowering therapy classes, number of current prescribed glucose-lowering therapy classes, a history of frailty (a proxy defined with a history of falls or lower limb fractures), previous prescription of blood pressure medication, previous prescription of statins (any prescription before treatment start) and history of metformin discontinuation within three months of initiation. Descriptive analysis We estimated the proportion of discontinuations for each drug class stratified into subgroups defined by the predictor features. We described the proportion of individuals discontinuing each drug class by subgroups. Treatment selection model development A Bayesian additive regression trees (BART) model was developed using the candidate features to predict discontinuation risk for each of the five glucose-lowering therapies. BART is a nonparametric regression modelling framework that captures predictor relationships and interactions without the need for pre-specification of exactly what form these relationships will take. This framework provides a lot of flexibility compared to more conventional methods when investigating treatment effect heterogeneity between different drug classes [ 16 ]. We used a standard BART model fitted using MCMC and the bartMachine package [ 17 ] (version 1.3.4.1) in R [ 18 ] (version 4.3.2). As we aimed to explore the potential for accurately predicting discontinuation, the whole cohort was used for model development and validation without resampling (or splitting the cohort into development and hold-out data). Individuals were excluded from model development if they had any missing features, as earlier work demonstrated that imputing missing values under a missing-completely-at-random assumption for a dataset of this magnitude does not markedly improve precision [ 19 ]. The discriminative power of the prediction model was assessed using the area under the receiver operating characteristic (AUROC) curves and calibration curves to ensure the predicted probability matched the observed proportions of discontinuation, overall and for each drug class. As a sensitivity analysis, separate models were also developed for discontinuation at 6- and 12-months. Potential heterogeneous treatment effects were evaluated with conditional average treatment effects (CATE) using the concordant–discordant approach previously proposed [ 5 , 6 ]. The CATE for an individual is conditional on clinical features and represents the predicted differential effects of two drug classes on the risk of discontinuation. In this approach, the cohort is split into subgroups based on predicted CATE estimates (defined by deciles), and the average CATE estimate within each subgroup is compared to estimates of average treatment effects derived from a set of alternative models fitted to each of the subgroups in turn. These latter models target the average treatment effect (ATE) within a population of individuals (rather than CATE), with desirable properties justified in the literature [ 20 ]. Alignment between both estimates provides evidence that ATEs are consistent across different inference methods within each subgroup. We used logistic regression as the primary approach for calculating ATEs within subgroups, estimating ATE as the difference in the risk of discontinuation between individuals receiving each drug class, adjusting for the full covariate set. Confidence intervals (CI) are estimated by refitting the models in bootstrapped datasets. We followed TRIPOD+AI reporting guidance (ESM Materials) [ 21 ]. Results We included 135,410 people with T2D initiating glucose-lowering therapies (182,194 initiations – 16,347 GLP-1RA, 71,460 DPP4i, 50,510 SGLT2i, 5,081 TZD and 38,796 SU) ( Table 1 , sFig. 1). Baseline clinical features by initiated drug class are reported in Table 1 . Of all treatment initiations, 26,138 (14.3%) were discontinued within the first three months since initiation. Discontinuation was highest for TZD initiations (19.6%) and lowest for DPP4i initiations (12.8%) (14.4% GLP-1RA, 14.9% SGLT2i, 15.7% SU). Discontinuation was higher at 6-months (range 17.2%-26.2%) and 12-months (23.2%-34.8%), but the relative proportions discontinuing each of the five drug classes was similar ( Figure 1 , sTable 1). View this table: View inline View popup Table 1: Demographic and clinical features of patients initiating the five glucose-lowering drug classes. All features were evaluated as potential predictors of discontinuation in model development. Data are mean [SD] and number (%). *Closest values to treatment start in the previous 6-months. A recorded baseline HbA 1c was a study inclusion criteria. Download figure Open in new tab Fig. 1: Proportion of individuals with type 2 diabetes discontinuing the five drug classes, at 3-months (primary outcome), 6-months, and 12-months. See sTable 1 for a breakdown of patient numbers for each therapy at each time point. 3-month discontinuation varied across the five therapy classes by age and routine clinical features Figure 2 shows the proportions of patients discontinuing therapy within 3-months stratified by sociodemographic and clinical features. When assessing sociodemographic features, there was some evidence that age at treatment initiation was associated with differences in discontinuation by drug, with higher discontinuation of GLP-1RA and SGLT2i in older adults, which was not observed for the other therapies. Discontinuation was higher in females and individuals of South Asian and Black ethnicity compared to White ethnicity, with no evidence of variation across the five therapy classes. Discontinuation was consistent by deprivation and smoking status. In terms of clinical features, we found higher discontinuation of GLP-1RA and SGLT2i in patients with longer duration of diabetes, lower discontinuation of SGLT2i, TZD and SU as the number of other current glucose-lowering therapies increased and higher discontinuation of SGLT2i, DPP4i and TZD as the number of previously prescribed therapies increased. Higher discontinuation was seen for patients not taking statins or blood pressure medication and those who have discontinued metformin in the past (similar across the five drug classes). When assessing laboratory measures, there was evidence of higher discontinuation at higher levels of HbA 1c for DPP4i and TZD and higher discontinuation at lower levels of BMI for GLP-1RA, TZD and SU ( Fig. 2 , sTable 2). As a sensitivity analysis, the 6- and 12-months discontinuation patterns across all features were similar to those seen for 3-month discontinuation (sFig. 2). Download figure Open in new tab Fig. 2: Proportion of individuals with T2D discontinuing the five drug classes at 3-months, stratified by sociodemographic and clinical features. A breakdown of patient numbers can be found in Table 1 and sTable 2. Discontinuation model development and validation The discontinuation model was developed using the whole cohort (n=162,779) with valid 3-month discontinuation outcome data and complete predictor features. As sensitivity analysis, model development was repeated for 6-months (n=157,327) and 12-months (n=142,610) discontinuation outcomes. The BART model converged quickly, and we ran 25,000 iterations, with the first 15,000 discarded as burn-in (trace plots are available on request). As would be anticipated for a within-sample evaluation, the calibration of the tolerability model was good, with the predicted discontinuation matching the observed discontinuation overall and per-therapy ( Fig. 3 ) and ranging from 8.8% to 23.5% across deciles of predicted discontinuation probabilities. Calibration was similarly good when predicting discrimination of individual therapy classes and for the 6- and 12-months discontinuation models (sFig. 3). However, model discrimination was weak. The 3-month discontinuation outcome model achieved an overall AUROC value of 0.61 (95%CI 0.61;0.62) ( Fig. 3 ), suggesting low utility to predict discontinuation accurately. Download figure Open in new tab Fig. 3: Calibration (1) and discrimination (2) plots for the model predicting discontinuation at 3-months, for overall (A) and per therapy (B) discontinuation. (1) Calibration of predicted and observed discontinuation for groups defined by decile of predicted discontinuation risk. Black line represents perfect calibration. (2) Receiver operating characteristic curve showing the discrimination between discontinuation and non-discontinuation. AUROC = area under the receiver operating characteristic curve. Discrimination was similarly low when assessing individual therapy classes (AUROC range 0.60–0.61, Fig.3 , sTable 3). Weak discrimination was also found for the 6-months (AUROC 0.61 [95%CI 0.61;0.62]) and 12-months (AUROC 0.61 [95%CI 0.61;0.62]) outcome models (sFig. 4, sTable 3). The mean CATE between therapies combinations ranged from 0.4% – 5.6 and the model demonstrated some heterogeneity in the predicted CATE estimates (sFig. 5). However, calibration by decile of model-predicted CATE estimates often did not align with estimates of ATE (sFig. 6). A similar finding was seen for 6- and 12-months discontinuation models (sFig. 7-8). Discussion In this large population-based study, we found that, although sociodemographic and clinical features are associated with differences in short-term discontinuation among new users of five major T2D drug classes (SGLT2i, GLP-1RA, DPP4i, TZD and SU), a model combining clinical features to predict differences in discontinuation risk by therapy for individual patients had low predictive ability (AUROC 0.61) and showed poor calibration of heterogeneous treatment effects. This finding suggests a limited potential for a precision medicine approach to T2D treatment based on short-term drug tolerability. The predictors with the largest differences in discontinuation across the five drug classes were age and diabetes duration, with higher discontinuation seen amongst older patients and, for SGLT2i and GLP-1RA, longer diabetes duration. Other features generally showed modest differences in short-term discontinuation that were broadly consistent by drug class, including higher discontinuation for patients who had previously discontinued metformin, females and patients of South-Asian/Black ethnicities, as well as lower discontinuation for patients currently taking statins and blood pressure medication. However, we showed that combining these differences into a prediction model was, even within sample, insufficient to predict discontinuation at the individual patient level. Although previous studies have examined features associated with discontinuation for individual therapy classes [7, 22, 23, 24, 25], our study is the first to describe clinical features associated with discontinuation considering all five therapy classes and the first to evaluate the utility of predicting treatment discontinuation at the individual patient level. Estimates of discontinuation at 12-months are similar to previous UK studies [7, 8, 9, 26, 27]. This is likely to reflect the healthcare system in the UK, as shown by differences with data from other countries, such as higher discontinuation in the CHOICE study (six European countries) [ 22 ] and lower discontinuation in a Danish cohort [ 24 ]. Additionally, differences in discontinuation levels between studies could be attributed to the low numbers of patients by drug class and differences in features (more males, lower duration of diabetes and lower HbA 1c ) [ 24 ]. Previous studies examining factors associated with discontinuation have, in keeping with our analysis, identified higher discontinuation at older ages and lower BMI [ 23 ], and higher discontinuation for patients initiating SGLT2i with lower eGFR and frailty [ 9 ]. Our study has limitations. Although our outcome definition was informed by previous studies [ 8 , 27 ], our dataset from routine clinical care did not allow us to capture the true underlying reason for discontinuation. Hence, we cannot be sure whether it represents poor tolerability or other factors. An alternative definition could have included a dose reduction (which may relate to a lack of tolerability); however, robust dosage data was unavailable in our dataset. Similarly, a diagnosis of adverse events leading to discontinuation could be used as part of the outcome definition; however, there are difficulties associated with differentiating between treatment-limiting and non-limiting adverse events in primary care data [ 28 ]. However, sensitivity analysis suggests our findings on overall discontinuation levels are consistent with previous UK studies [ 7 , 8 , 9 ], supporting our choice. The requirement for a gap in prescriptions above 180 days means we can reasonably ensure the treatment has been discontinued for a sustained period. However, it does not fully distinguish between temporary and permanent discontinuation [ 29 ]. Reassuringly, all our analyses were consistent when assessing discontinuation at 6- and 12-months, supporting the generalisability of our outcome definition. This study deliberately only looked at baseline features as we aimed to establish who might discontinue when initiating a new drug to potentially inform baseline treatment decisions, and we did not intend to look at what might affect discontinuation after treatment has started, unlike previous work [ 30 ]. This study contains lower numbers of TZD initiations versus other therapies, but compared to other studies, this study includes higher numbers of therapy initiations for all therapies [ 7 , 8 ]. Finally, as our analysis was performed using UK data, findings may not be generalisable to other countries and health systems. This is the first study to explore differential discontinuation between all five major glucose-lowering therapies using routine clinical features, providing notable evidence of higher discontinuation for older adults starting GLP-1RA and SGLT2i. Importantly, we showed that routine clinical features do not provide enough clinical utility alone to inform treatment targeting by predicting patients at high risk of short-term discontinuation of glucose-lowering therapy. The model developed in this study could be deployed in conjunction with other treatment selection tools [ 5 , 6 , 11 ] to provide further differentiation in situations of equipoise. Despite this, our analysis does not undermine the importance of robust studies to understand susceptibility to drug-specific side-effects, which could include genetics and multi-omics [ 31 , 32 ]. Considering treatment benefits, recent work has identified practical applications of routine clinical features to target T2D medication based on glucose-lowering, with robust heterogenous treatment effects for all five drug classes, providing clear evidence to guide choice of therapy for individual patients [ 5 , 6 , 11 , 33 , 34 , 35 , 36 ]. Conclusion Our study demonstrates that short-term discontinuation varies modestly across the five major T2D drug classes after metformin by certain clinical features, including age and diabetes duration. Despite this variation, our analysis suggests a limited scope to predict the likelihood of short-term tolerability by drug class for individual patients. This indicates that T2D clinical prescribing decisions should focus on drug-specific glycaemic and cardiometabolic benefits, alongside specific side-effect risks, instead of potential tolerability. Data Availability The UK routine clinical data analysed during the current study are available in the CPRD repository (CPRD; https://cprd.com/research-applications), but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. For re-using these data, an application must be made directly to CPRD. All R code used for the analysis is provided at https://github.com/Exeter-Diabetes/CPRD-Pedro-T2DDiscontinuation. Funding and Assistance The authors acknowledge support from the Medical Research Council (UK) (MR/N00633X/1), and the National Institute for Health and Care Research Exeter Biomedical Research Centre. JMD is supported by a Wellcome Trust Early Career award (227070/Z/23/Z). ATH and BMS are supported by the NIHR Exeter Clinical Research Facility. The views expressed are those of the authors and not necessarily those of the NHS, the NIHR or the Department of Health. Conflict of Interest All authors declare that there are no relationships or activities that might bias, or be perceived to bias, their work. Author Contributions PC, JMD, BMS, ATH conceived and designed the study. PC analysed the data and developed the code. KGY and RH helped with curating the CPRD dataset. All authors contributed to the writing of the article, provided support for the analysis and interpretation of results, critically revised the article and approved the final article. Data & Code The UK routine clinical data analysed during the current study are available in the CPRD repository (CPRD; https://cprd.com/research-applications ), but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. For re-using these data, an application must be made directly to CPRD. All R code used for the analysis is provided at https://github.com/Exeter-Diabetes/CPRD-Pedro-T2DDiscontinuation . Acknowledgements For the purpose of open access, the author has applied a CC BY public copyright license to any Author Accepted Manuscript version arising from this submission. This article is based in part on data from the CPRD obtained under license from the UK Medicines and Healthcare products Regulatory Agency. CPRD data are provided by patients and collected by the UK National Health Service (NHS) as part of their care and support. Approval for CPRD data access and the study protocol was granted by the CPRD Independent Scientific Advisory Committee (eRAP protocol number: 22_002000). The authors acknowledge contributions from the wider MASTERMIND consortium who supported this work (see ESM for a list of MASTERMIND consortium members). Footnotes Proposed journal: Diabetologia (FOLLOWED BY) Diabetes, Obesity and Metabolism References [1]. ↵ Z. Bin , A. Rayner , E. Gregg and et al. , “ Worldwide trends in diabetes prevalence and treatment from 1990 to 2022: a pooled analysis of 1108 population-representative studies with 141 million participants ,” The Lancet , vol. 404 , no. 10467 , pp. 2077 – 2093 , 2024 . OpenUrl [2]. ↵ A. Kaiser , N. Zhang and W. Pluijm , “ Global prevalence of type 2 diabetes over the next ten years (2018-2028) ,” Diabetes , vol. 67 , pp. 202 – LB , 2018 . OpenUrl CrossRef [3]. ↵ N. Cho , J. Shaw , S. Karuranga , Y. Huang , J. da Rocha Fernandes , A. Ohlrogge and B. Malanda , “ IDF Diabetes Atlas: Global estimates of diabetes prevalence for 2017 and projections for 2045 ,” Diabetes Research and Clinical Practice , vol. 138 , pp. 271 – 281 , 2018 . OpenUrl CrossRef PubMed [4]. ↵ M. Davies , V. Aroda , B. Collins , R. Gabbay , J. Green , N. Maruthur , S. Rosas , S. Prato , C. Mathieu , G. Mingrone , P. Rossing , T. Tankova , A. Tsapas and J. Buse , “Management of Hyperglycemia in Type 2 Diabetes, 2022. A Consensus Report by the American Diabetes Association (ADA) and the European Association for the Study of Diabetes (EASD) ,” Diabetes Care , vol. 45 , no. 11 , pp. 2753 – 2786 , 2022 . OpenUrl CrossRef PubMed [5]. ↵ P. Cardoso , K. Young , A. Nair , R. Hopkins , A. McGovern , E. Hairder , P. Karunaratne , L. Donnelly , B. Mateen , N. Sattar , R. Holman , J. Bowden , A. Hattersley , E. Pearson , A. Jones , B. Shields , T. McKinley and J. Dennis , “ Phenotype-based targeted treatment of SGLT2 inhibitors and GLP-1 receptor agonists in type 2 diabetes ,” Diabetologia , vol. 67 , pp. 822 – 836 , 2024 . OpenUrl PubMed [6]. ↵ J. Dennis , K. Young , A. McGovern , B. Mateen , S. Vollmer , M. Simpson , W. Henley , R. Holman , N. Sattar , E. Pearson , A. Hattersley , A. Jones and B. Shields , “ Development of a treatment selection algorithm for SGLT2 and DPP-4 inhibitor therapies in people with type 2 diabetes: a retrospective cohort study ,” The Lancet Digital Health , vol. 4 , no. 12 , pp. e873 – e883 , 2022 . OpenUrl [7]. ↵ T. Wilke , S. Mueller , A. Groth , B. Berg , A. Fuchs , M. Sikirica , J. Logie , A. Martin and U. Maywald , “ Non-Persistence and Non-Adherence of Patients with Type 2 Diabetes Mellitus in Therapy with GLP-1 Receptor Agonists: A Retrospective Analysis ,” Diabetes Therapy , vol. 7 , pp. 105 – 124 , 2015 . OpenUrl PubMed [8]. ↵ J. Wilding , T. Godec , K. Khunti , S. Pocock , R. Fox , L. Smeeth , P. Clauson , P. Fenici , N. Hammar and J. Medina , “ Changes in HbA1c and weight, and treatment persistence, over the 18 months following initiation of second-line therapy in patients with type 2 diabetes: results from the United Kingdom Clinical Practice Research Datalink ,” BMC Medicine , vol. 16 , p. 116 , 2018 . OpenUrl PubMed [9]. ↵ W. Alkabbani , B. Shah , A. Zongo , D. Eurich , W. Alsabbagh and J. Gamble , “ Post-initiation predictors of discontinuation of the sodium-glucose cotransporter-2 inhibitors: A comparative cohort study from the United Kingdom ,” Diabetes, Obesity and Metabolism , vol. 25 , no. 12 , pp. 3490 – 3500 , 2023 . OpenUrl PubMed [10]. ↵ L. Razavi-Nematollahi and F. Ismail-Beigi , “ Adverse Eaects of Glycemia-Lowering Medications in Type 2 Diabetes ,” Current Diabetes Reports , vol. 19 , p. 132 , 2019 . OpenUrl PubMed [11]. ↵ J. Dennis , K. Young , P. Cardoso , L. Güdemann , A. McGovern , A. Farmer , R. Holman , N. Sattar , T. McKinley , E. Pearson , A. Jones , B. Shields and A. Hattersley , “ A five-drug class model using routinely available clinical features to optimise prescribing in type 2 diabetes: a prediction model development and validation study ,” The Lancet , 2025 . [12]. ↵ A. McGovern , M. Hogg , B. Shields , N. Sattar , R. Holman , E. Pearson , A. Hattersley , G. Angus and J. Dennis , “ Risk factors for genital infections in people initiating SGLT2 inhibitors and their impact on discontinuation ,” Emerging Technologies, Pharmacology and Therapeutics , vol. 8 , no. e001238 , 2020 . [13]. ↵ W. Polonsky and R. Henry , “ Poor medication adherence in type 2 diabetes: recognizing the scope of the problem and its key contributors ,” Patient Preference and Adherence , vol. 10 , pp. 1299 – 1307 , 2016 . OpenUrl [14]. ↵ A. Wolf , D. Dedman , J. Campbell , H. Booth , D. Lunn , J. Chapman and P. Myles , “ Data resource profile: Clinical Practice Research Datalink (CPRD) Aurum ,” International Journal of Epidemiology , vol. 48 , no. 6 , pp. 1740 – 1740g , 2019 . OpenUrl CrossRef PubMed [15]. ↵ L. Rodgers , M. Weedon , W. Henley , A. Hattersley and B. Shields , “ Cohort profile for the MASTERMIND study: using the Clinical Practice Research Datalink (CPRD) to investigate stratification of response to treatment in patients with type 2 diabetes ,” BMJ Open , vol. 7 , p. e017989 , 2017 . OpenUrl Abstract / FREE Full Text [16]. ↵ H. Chipman , E. George and R. McCulloch , “ BART: Bayesian additive regression trees ,” The annals of applied statistics , vol. 4 , no. 1 , pp. 266 – 298 , 2010 . OpenUrl [17]. ↵ A. Kapelner and J. Bleich , “ bartMachine: machine learning with Bayesian additive regression trees ,” Journal of Statistical Software , vol. 70 , no. 4 , pp. 1 – 40 , 2016 . OpenUrl [18]. ↵ R Core Team , “R: a language and environment for statistical computing,” 2021 . [Online]. Available: https://www.R-project.org/ . [19]. ↵ P. Cardoso , J. Dennis , J. Bowden , B. Shields and T. McKinley , “ Dirichlet process mixture models to impute missing predictor data in counterfactual prediction models: an application to predict optimal type 2 diabetes therapy ,” BMC Medical Informatics and Decision Making , vol. 24 , no. 12 , 2024 . [20]. ↵ F. E. Harrell , Regression modeling strategies , Springer International Publishing , 2016 . [21]. ↵ G. Collins , K. Moons , P. Dhiman and et al. , “ TRIPOD + AI statement: updated guidance for reporting clinica prediction models that use regression or machine learning methods ,” BMJ , no. 385 , p. e078378 , 2024 . [22]. ↵ C. Östenson , S. Matthaei , M. Reaney , T. Krarup , B. Guerci , J. Kiljanski , C. Salaun-Martin , H. Sapin , D. Bruhn , C. Mathieu and M. Theodorakis , “ Treatment outcomes after initiation of exenatide twice daily or insulin in clinical practice: 12-month results from CHOICE in six European countries ,” Diabetes, Metabolic Syndrome and Obesity , vol. 6 , pp. 171 – 185 , 2013 . OpenUrl [23]. ↵ T. Weiss , L. Yang , R. Carr , S. Pal , B. Sawhney , R. Boggs , S. Rajpathak and K. Iglay , “ Real-world weight change, adherence, and discontinuation among patients with type 2 diabetes initiating glucagon-like peptide-1 receptor agonists in the UK ,” BMJ Open Diabetes Research and Care , vol. 10 , p. e002517 , 2022 . OpenUrl Abstract / FREE Full Text [24]. ↵ M. Lassen , N. Johansen , D. Modin , A. Catarig , B. Vistisen , H. Amadid , E. Zimmermann , G. Gislason and T. Biering-Sørensen , “ Adherence to glucagon-like peptide-1 receptor agonist treatment in type 2 diabetes mellitus: A nationwide registry study ,” Diabetes, Obesity and Metabolism , vol. 26 , no. 11 , pp. 5239 – 5250 , 2024 . OpenUrl PubMed [25]. A. Plat , F. Beest , S. Kessabi , M. Groot and R. Herings , “ Change of initial oral antidiabetic therapy in type 2 diabetic patients ,” Pharmacy World & Science , vol. 31 , pp. 622 – 626 , 2009 . OpenUrl PubMed [26]. A. Maguire , B. Mitchell and J. Ruzafa , “ Antihyperglycaemic treatment patterns, observed glycaemic control and determinants of treatment change among patients with type 2 diabetes in the United Kingdom primary care: a retrospective cohort study ,” BMC Endocrine Disorders , vol. 14 , p. 73 , 2014 . OpenUrl PubMed [27]. ↵ E. Lamberts , G. Nijpels , L. Welschen , J. Hugtenburg , J. Dekker , P. Souverein and M. Bouvy , “ Long term patterns of use after initiation of oral antidiabetic drug therapy ,” Pharmacoepidemiology and Drug Safety , vol. 20 , no. 4 , pp. 351 – 358 , 2010 . OpenUrl PubMed [28]. ↵ A. Climens , E. Pain , A. Boss and A. Shaunik , “ Understanding reasons for treatment discontinuation, attitudes and education needs among people who discontinue types 2 diabetes treatment: results from an online patient survey in the USA and UK ,” Diabetes Therapy , vol. 11 , pp. 1873 – 1881 , 2020 . OpenUrl PubMed [29]. ↵ Y. Tang , T. Weiss , J. Liu , S. Rajpathak and K. Khunti , “ Metformin adherence and discontinuation among patients with type 2 diabetes: A retrospective cohort study ,” Journal of Clinical & Translational Endocrinology , vol. 20 , p. 100225 , 2020 . OpenUrl PubMed [30]. ↵ A. McGovern , M. Hogg , B. Shields , N. Sattar , R. Holman , E. Pearson , A. Hattersley , A. Jones and J. Dennis , “ Risk factors for genital infections in people initiating SGLT2 inhibitors and their impact on discontinuation ,” BMJ Open Diabetes Research & Care , vol. 8 , p. e001238 , 2020 . OpenUrl [31]. ↵ L. Güdemann , K. Young , N. Thomas , R. Hopkins , R. Challen , A. Jones , A. Hattersley , E. Pearson , B. Shields , J. Bowden , J. Dennis and A. McGovern , “ Safety and eaectiveness of SGLT2 inhibitors in a UK population with type 2 diabetes and aged over 70 years: an instrumental variable approach ,” Diabetologia , vol. 67 , pp. 1817 – 1827 , 2024 . OpenUrl PubMed [32]. ↵ E. Pearson , “ Diabetes: is there a future for pharmacogenomics guided treatment? ,” Clinical Pharmacology & Therapeutics , vol. 106 , no. 2 , pp. 329 – 337 , 2019 . OpenUrl PubMed [33]. ↵ J. Dennis , “ Precision medicine in type 2 diabetes: using individualised prediction models to optimize selection of treatment ,” Diabetes , vol. 69 , no. 10 , pp. 2075 – 2085 , 2020 . OpenUrl CrossRef PubMed [34]. ↵ D. Cherney , M. Cooper , I. Tikkanen , E. Pfarr , O. Johansen , H. Woerle , U. Broedl and S. Lund , “ Pooled analysis of Phase III trials indicate constrasting influences of renal function on blood pressure, body weight, and HbA1c reductions with empagliflozin ,” Kidney International , vol. 93 , no. 1 , pp. 231 – 244 , 2018 . OpenUrl CrossRef PubMed [35]. ↵ J. Dennis , W. Henley , M. Weedon , M. Lonergan , L. Rodgers , A. Jones , W. Hamilton , N. Sattar , S. Janmohamed , R. Holman , E. Pearson , B. Shields and A. Hattersley , “ Sex and BMI alter the benefits and risks of sulfonylureas and thiazolidinediones in type 2 diabetes: a framework for evaluating stratification using routine clinical and individual trial data ,” Diabetes Care , vol. 41 , no. 9 , pp. 1844 – 1853 , 2018 . OpenUrl Abstract / FREE Full Text [36]. ↵ B. Shields , J. Dennis , C. Angwin , F. Warren , W. Henley , A. Farmer , N. Sattar , R. Holman , A. Jones , E. Pearson and A. Hattersley , “ Patient stratification for determining optimal second-line and third-line therapy for type 2 diabetes: the TriMaster study ,” Nature Medicine , vol. 29 , no. 2 , pp. 376 – 383 , 2023 . OpenUrl PubMed View the discussion thread. Back to top Previous Next Posted February 20, 2025. Download PDF Supplementary Material Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Evaluating prediction of short-term tolerability of five type 2 diabetes drug classes using routine clinical features: UK population-based study Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Evaluating prediction of short-term tolerability of five type 2 diabetes drug classes using routine clinical features: UK population-based study Pedro Cardoso , Katie G. Young , Rhian Hopkins , Bilal A. Mateen , Ewan R. Pearson , Andrew T. Hattersley , Trevelyan J. McKinley , Beverley M. Shields , John M. Dennis medRxiv 2025.02.17.25322406; doi: https://doi.org/10.1101/2025.02.17.25322406 Share This Article: Copy Citation Tools Evaluating prediction of short-term tolerability of five type 2 diabetes drug classes using routine clinical features: UK population-based study Pedro Cardoso , Katie G. Young , Rhian Hopkins , Bilal A. Mateen , Ewan R. Pearson , Andrew T. Hattersley , Trevelyan J. McKinley , Beverley M. Shields , John M. Dennis medRxiv 2025.02.17.25322406; doi: https://doi.org/10.1101/2025.02.17.25322406 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Epidemiology Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (300) Cardiovascular Medicine (4438) Dentistry and Oral Medicine (444) Dermatology (383) Emergency Medicine (608) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1509) Epidemiology (15229) Forensic Medicine (30) Gastroenterology (1126) Genetic and Genomic Medicine (6600) Geriatric Medicine (668) Health Economics (998) Health Informatics (4538) Health Policy (1369) Health Systems and Quality Improvement (1613) Hematology (542) HIV/AIDS (1264) Infectious Diseases (except HIV/AIDS) (15919) Intensive Care and Critical Care Medicine (1103) Medical Education (623) Medical Ethics (147) Nephrology (667) Neurology (6600) Nursing (346) Nutrition (998) Obstetrics and Gynecology (1145) Occupational and Environmental Health (957) Oncology (3333) Ophthalmology (974) Orthopedics (369) Otolaryngology (420) Pain Medicine (436) Palliative Medicine (130) Pathology (663) Pediatrics (1693) Pharmacology and Therapeutics (692) Primary Care Research (711) Psychiatry and Clinical Psychology (5447) Public and Global Health (9233) Radiology and Imaging (2199) Rehabilitation Medicine and Physical Therapy (1370) Respiratory Medicine (1196) Rheumatology (593) Sexual and Reproductive Health (712) Sports Medicine (530) Surgery (712) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a0109443ebbd0db4',t:'MTc3OTY3MDU3NA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();