Extracting Dosage Per Day From Free-Text Medication Prescriptions

VA L U E I N H E A LT H 1 9 ( 2 0 1 6 ) A 3 4 7 – A 7 6 6

indirect comparison (MAIC), and simulated treatment comparison (STC).  Methods: Patient-level datasets for two oncology trials were simulated. The first trial (IPDaccessible trial) comprised Treatment A and Placebo patients; the second trial (APD-accessible trial) comprised Treatment B and Placebo patients. The outcome of interest was progression-free survival. Patient characteristics of interest were age, gender, disease stage, laboratory test positivity, and brain metastases history. Each method (CCMM, MAIC, and STC) was implemented to obtain an adjusted Treatment A vs. Treatment B hazard ratio (HR). The naïve Treatment A vs. Treatment B HR was calculated using Bucher’s method as an unadjusted reference point, while data from a head-to-head propensity score-adjusted simulation of Treatment A vs. Treatment B were used as a fully adjusted reference point.  Results: Before adjustment, the Treatment A vs. Placebo HR was 0.24, and the Treatment B vs. Placebo HR was 0.45, yielding a naïve Treatment A vs. Treatment B HR of 0.55 (95% CI: 0.42, 0.72). The Treatment A vs. Treatment B HR after adjustment was 0.62 (95% CI: 0.47, 0.83) upon applying CCMM, 0.69 (95% CI: 0.53, 0.89) upon applying MAIC, and 0.69 (95% CI: 0.56, 0.84) upon applying STC. Data from the head-to-head propensity score-adjusted simulation yielded an HR of 0.79 (95% CI: 0.63, 0.99).  Conclusions: In this simulation, MAIC and STC performed similarly in adjusting for cross-trial differences while CCMM yielded the least amount of adjustment. Each method has advantages and disadvantages, and more research is needed to understand the performance of each method in specific situations. PRM195 The Added Value of Real-World Evidence to Connect Disconnected Networks for Network Meta-Analysis: A Case Study in Rheumatoid Arthritis Jenkins DA, Martina R, Dequen P, Bujkiewicz S, Abrams K University of Leicester, Leicester, UK

Objectives: There are many circumstances under which networks of evidence may be ‘disconnected’ and network meta-analysis (NMA) cannot be conducted, unless additional assumptions are made. However, real-world evidence (RWE) which is becoming a more widely used source of clinical data to complement randomised evidence for relative effectiveness assessment, could help inform missing ‘connections’ within a network. We consider the impact of RWE on NMA to compare existing biologic DMARDS in rheumatoid arthritis (RA) for second-line therapy in a disconnected network.  Methods: A literature search was undertaken to identify RCTs evaluating second-line biological therapies in RA. Patient data from two European registries were also accessed. Standard Bayesian NMA and naïve pooling of standard of care were applied and evaluated. Alternatively, RWE and RCT data were combined in an NMA to connect the RA network of studies.  Results: Only 4 RCTs were identified for second-line biologics with one treatment (Golimumab) disconnected from the network. All methods applied were effective in allowing for the comparison of Golimumab against all other treatments. For example, Golimumab had increased probability of achieving remission by 7.6% (CI: 2.3% to 13.6%) compared to standard of care, an estimate that would not have been possible to obtain if using RCT data alone. The addition of RWE to the RCT data led to a decrease in the level of uncertainty of the probability of remission in Rituximab compared to standard of care from 8.3% (CI: 4.9% to 12%) to 7.2% (CI: 4.1% to 10.7%).  Conclusions: The use of RWE was a useful approach here. By bridging disconnected networks of RCT evidence, RWE allowed evaluation of treatment options otherwise not comparable via a standard NMA of RCTs alone. In addition, estimates of effect of treatments already included in the RCT network were obtained with higher precision when including the RWE. PRM196 Comparison of Bayesian Network Meta-Analyses in a Winbugs and SAS Framework Le Moine J, Abeysinghe SS RTI Health Solutions, Didsbury, UK

Objectives: While several types of software are available, WinBUGS is the preferred statistical package to conduct network meta-analysis (NMA). Despite being widely used in the pharmaceutical industry, SAS use in NMA is limited. This research aims to compare results from meta-analyses conducted in WinBUGs and SAS.  Methods: NMA of trial data were conducted in WinBUGS and SAS and their results compared. Networks of evidence with different levels of complexity, in terms of network structure, and the number of treatments and studies, were considered. Fixed (FE) and random effects (RE) analyses were conducted.  Results: Analyses of a head-to-head network (TSD2 2011, Beta-blockers example: 2 treatments, 22 studies) showed strong consistency between SAS and WinBUGS for FE (log-OR [95%CI]: -0.2614 [-0.3582;-0.1623] vs -0.2622 [-0.36;-0.1637] respectively for SAS and WinBUGS) and RE analyses (-0.2488 [-0.3746;-0.1168] vs -0.25 [-0.759;-0.1164] respectively). A closed-loop network (Jones et al. 2011, Cirrhosis example: 3 treatments, 26 studies) showed more consistency with FE than with RE. Observed differences between the means ranged from 0.0001 to 0.006 for the FE model, and from 0.01 to 0.16 for the RE model. All significant differences were consistent between WinBUGS and SAS results. Similar results were observed with a star-shaped network (TSD3 2011, Rheumatoid arthritis example: 7 treatments, 12 studies). SAS and WinBUGS produced consistent estimates of mean and standard deviation. However, in the RE analysis, credible intervals led to different conclusions. Etanercept was found to be significantly superior to methotrexate in WinBUGS but not in SAS (95% CI: [0.3717;7.391] vs [0.0165;7.0553]). The significant WinBUGS result was consistent with the head-to-head evidence.  Conclusions: SAS provides an alternative to WinBugs to conduct NMA, and constitutes a potential means to validate WinBUGS results. However its results can significantly differ from WinBUGS depending on the nature of the network and so should be used with caution. PRM197 Adjusting Variance Parameters to Incorporate Uncertainty into Health Economics Models Following Treatment Switching

A393

Ray J1, Bennett I2, Paracha N2 1F. Hoffmann-La Roche, Basel, Switzerland, 2F. Hoffman-La Roche, Basel, Switzerland

Objectives: Trials designed to estimate the efficacy of new oncological treatments commonly permit patients switch from standard care (SC) to active treatment. Clinical regulators focus on the safety and efficacy of multiple endpoints, whereas clear demonstration of an overall survival (OS) gain is prioritized by many HTA agencies. When patients randomized to receive SC subsequently receive active treatment following disease progression, OS gains are confounded. Several statistical methods have been proposed to adjust for the impact of patients switching treatment to estimate counterfactual survival, one of which is the RPSFT method. Health economic models using these adjusted survival times may fail to introduce the additional uncertainty following the application of this method when implementing parametric survival analysis, increasing the likelihood of an incorrect decision.  Methods: Twenty datasets of 400 patients each were simulated assuming a Weibull distribution, where 70% of patients initially treated with SC switched to receive the active treatment following progression. Two scenarios were compared. Initially, Weibull survival estimates and covariance parameters were estimated using the RPSFT-adjusted survival times. These were compared with a similar approach with the addition of multiplying the components of the covariance matrix that involve uncertainty around the SC treatment parameter by the adjustment factor identified in the RPSFT statistical model for each individual simulated dataset.  Results: Mean undiscounted incremental life expectancy between the two scenarios were compared. Once the covariance matrices were sampled 5000 times using PSA, inflated confidence intervals in the second scenario suggested that failing to incorporate the additional uncertainty could lead to incorrect decisions. The probability of the new treatment being considered less efficacious compared to SC was 0.6% compared to 5.68% with appropriately accounting for the increased uncertainty.  Conclusions: Failing to account for uncertainty when applying treatment switching methods in health economic models could misinform decisions determining patients’ access to treatments. PRM198 Extracting Dosage Per Day From Free-Text Medication Prescriptions TornblomTornblomTornblomTörnblom M1, Bergman G2, Jørgensen L3, Fackle-Fornius E1, Rosenlund M2 University, Stockholm, Sweden, 2IMS Health, Solna, Sweden, 3Pygargus/IMS Health, Solna, Sweden 1Stockholm

Objectives: The Swedish prescribed drug register contains dose instructions as written by the physician. A challenge is to convert the text into a number of doses per day which can be used to calculate for example duration of treatment. The objective of this study is to compare algorithms for named entity recognition to extract dosage per day.  Methods: Two sequence models, Hidden Markov Model (HMM) and Conditional Random Fields (CRF), were used to predict label sequences. The HMM and CRF were compared using different measures of prediction: precision, recall, F-score and accuracy. We also evaluated how prediction was effected by including more labels and features; for CRF models we used 12 labels for both models with 2 and 11 feature types respectively, for HMM models we used 12, 15 and 18 labels respectively. Using the predicted labels, a rule-based algorithm was used to predict dosage per day. Prediction of dosage per day was evaluated using accuracy.  Results: Label prediction: As expected, increasing the number of labels/ features increased the F-score. The CRF model with 11 feature types had a F-score of 0.989 compared to 0.972 using two feature types. The HMM model with 15 and 18 labels both achieved a F-score of 0.986 compared to 0.966 using 12 labels. In terms of precision and recall the performance of the CRF and HMM varied. Dosage prediction: The CRF model with 11 feature types achieved 97.2% accuracy. The HMM with 15 labels achieved a higher accuracy than with 18 labels (95.7% versus 95.5%).  Conclusions: The CRF has the highest accuracy in label and dosage per day prediction. The HMM model also has comparably high accuracy but is generally lower than the CRF. We recommend CRF over HMM for named entity recognition on prescription text; it is time efficient and predicts dosage per day with high accuracy. PRM199 Identifying Higher-Value Subpopulations For Treatment In Heterogenous Rare Diseases: An Example Study of Early Responders to Teduglutide for Short Bowel Syndrome Chen K1, Xie J2, Tang C2, Zhao J3, Olivier C4, Jeppesen PB5, Signorovitch J3 1Shire plc, Lexington, MA, USA, 2Analysis Group, New York, NY, USA, 3Analysis Group, Boston, MA, USA, 4Shire GmbH, Zug, Switzerland, 5Rigshospitalet, Copenhagen, Denmark

Objectives: Measurement of drug efficacy in rare diseases is often complicated by patient heterogeneity and limited sample sizes. Even when treatments show significant efficacy overall, small samples can limit understanding of subpopulation effects. We applied statistical methods for identifying higher-value subpopulations to a study of early teduglutide (TED) response in short bowel syndrome (SBS) with parenteral nutrition (PN) dependency.  Methods: SBS patients in a phase 3 trial (NCT00798967; EudraCT2008-006193-15) were randomized to TED (n= 43) or placebo (n= 43). Early response was measured as ≥ 20% PN volume reduction at weeks 20 and 24. Separate response prediction models were developed for each treatment using penalized regression and cross-validation to avoid over-fitting. Patients were ranked by predicted early response rate difference with TED vs. placebo. To measure the value of this ranking for identifying early responder subpopulations, observed response rate differences were estimated in subpopulations defined by the ranks.  Results: In the 24-week trial, TED was associated with a significantly higher early response rate (63%) vs. placebo (30%), and with an even higher 93.3% response rate in a 24-month extension study. Predictors of early response to TED included: older age, absence of ileocecal valves, lower percent of colon remaining, volvulus as the cause of intestine resection, higher baseline PN volume, and longer time since start of PN dependency. Higher percent of colon remaining and volvulus were predictors for early response to placebo. An early responder subpopulation consisting of the top-ranked 60% of patients had early response rates of 88% for