Background and context
Large databases of health information are an important resource to study the use and outcomes of health services including the use of medications (Cadarette & Wong, Reference Cadarette and Wong2015; Metge et al., Reference Metge, Grymonpre, Dahl and Yogendran2005; Murdoch & Detsky, Reference Murdoch and Detsky2013; Schneeweiss & Avorn, Reference Schneeweiss and Avorn2005; Zhan & Miller, Reference Zhan and Miller2003). Information on the prevalence, incidence, and duration of drug therapy is important in health research, health system planning, and assessment of appropriate prescribing for treatment patterns and burden (Galvin et al., Reference Galvin, Moriarty, Cousins, Cahir, Motterlini, Bradley, Hughes, Bennett, Smith, Fahey and Kenny2014; Moriarty et al., Reference Moriarty, Bennett, Fahey, Kenny and Cahir2015a, Reference Moriarty, Hardy, Bennett, Smith and Fahey2015b; Schneeweiss & Avorn, Reference Schneeweiss and Avorn2005). Moreover, as the global population of adults 65 years and older continues to grow, the need will also grow for timely and accurate information not only for prescribed medications but also for non-prescription medications and natural health product (NHP). Standardized coding and classification of medication data can improve the efficiency in data collection and curation processes, which are complex processes due to heterogeneous formats including generic names (e.g., acetaminophen), trade names (e.g., Tylenol), and numeric drug identifiers (e.g., 02046040) (Nikiema et al., Reference Nikiema, Liang, Després and Motulsky2021; Richesson, Reference Richesson2014).
The mapping of medication data to standardized terminologies such as the RxNorm ontology (RxNorm, n.d.) has been proposed to allow efficient analysis and interpretation of drug data (Nikiema et al., Reference Nikiema, Liang, Després and Motulsky2021; Richesson, Reference Richesson2014). The performance of this mapping to standardized terminologies has been evaluated with medication data from hospital pharmacy systems (Hernandez et al., Reference Hernandez, Podchiyska, Weber, Ferris and Lowe2009; Waters et al., Reference Waters, Malecki, Lail, Mak, Saha, Jung, Imrit, Razak and Verma2023), electronic health records (Zhou et al., Reference Zhou, Plasek, Mahoney, Chang, DiMaggio and Rocha2012), drug adverse events database (Veronin et al., Reference Veronin, Schumaker, Dixit, Dhake and Ogwo2020), multi-site clinical trial (Lockery et al., Reference Lockery, Rigby, Collyer, Stewart, Woods, McNeil, Reid and Ernst2019; Richesson et al., Reference Richesson, Smith, Malloy and Krischer2010), and longitudinal cohorts (Richesson et al., Reference Richesson, Smith, Malloy and Krischer2010). For prospective clinical studies, the ASPREE (Lockery et al., Reference Lockery, Rigby, Collyer, Stewart, Woods, McNeil, Reid and Ernst2019) clinical trial in older adults and the 45 and Up study (Gnjidic et al., Reference Gnjidic, Pearson, Hilmer, Basilakis, Schaffer, Blyth and Banks2015) reported a method of structured medication data collection based on a list of common medications with the option of free-text data entry for other medications. Both studies used a structured process of automated and manual coding for the curation of the free-text data by medication experts (Gnjidic et al., Reference Gnjidic, Pearson, Hilmer, Basilakis, Schaffer, Blyth and Banks2015; Lockery et al., Reference Lockery, Rigby, Collyer, Stewart, Woods, McNeil, Reid and Ernst2019). Systematic approaches for the curation of large free-text medication data have involved automated and manual approaches (Richesson, Reference Richesson2014; Veronin et al., Reference Veronin, Schumaker, Dixit, Dhake and Ogwo2020).
The Canadian Longitudinal Study on Aging (CLSA) is a population-based research platform established to better understand how biological, medical, psychological, and social determinants have an impact in maintaining health and in the development of disease and disability as people age (P. Raina et al., Reference Raina, Wolfson, Kirkland, Griffith, Balion, Cossette, Dionne, Hofer, Hogan, van den Heuvel, Liu-Ambrose, Menec, Mugford, Patterson, Payette, Richards, Shannon, Sheets, Taler and Young2019; P. S. Raina et al., Reference Raina, Wolfson, Kirkland, Griffith, Oremus, Patterson, Tuokko, Penning, Balion, Hogan, Wister, Payette, Shannon and Brazil2009). The complete documentation of all drugs and NHP used every 3 years over 20 years in a cohort of more than 30,000 participants requires efficient data mapping and curation processes. In this article, we describe a three-step process for the data entry and mapping of drug data to the Health Canada Drug Product Database (DPD) by CLSA interviewers, as well as the development and validation of a cleaning process of free-text/numeric drug and NHP inputs in a software algorithm approach followed by manual recoding.
Methods
Study population
The recruitment and baseline evaluations of the 51,338 CLSA participants aged 45–85 years at enrolment was completed in 2015 (P. Raina et al., Reference Raina, Wolfson, Kirkland, Griffith, Balion, Cossette, Dionne, Hofer, Hogan, van den Heuvel, Liu-Ambrose, Menec, Mugford, Patterson, Payette, Richards, Shannon, Sheets, Taler and Young2019). The complete CLSA cohort is composed of the Tracking cohort of 21,241 participants who provide data via telephone interviews and the Comprehensive cohort of 30,097 participants who provide data via in-person home interviews and visits to a data-collection site. Comprehensive participants provided data in English and French on all regularly used drug and NHPs.
Drug and NHP data collection/mapping drug data to Health Canada database
In the first of a three-step process, Drug and NHP data were entered in the CLSA data collection software by interviewers who were trained to identify the relevant information from medication packaging (Figure 1). During an in-home visit, CLSA interviewers asked participants to present all regularly scheduled or taken medications (i.e., scheduled, once a day, every other day, taken occasionally, and as required), including prescription, non-prescription, over-the-counter, herbals, vitamins, or NHPs in all routes of administration. Information on study drugs and drugs commercialized in other countries than Canada was also collected. The interviewer entered either the generic name (e.g., atorvastatin), trade name (e.g., Lipitor), or drug identification number (DIN) (e.g., 02230711) in a type-to-search box that mapped the drug input to the Health Canada DPD and generated a list of corresponding generic or trade drug names. In the absence of adequate drug name correspondence, the name/DIN was entered as a free-text/numeric input. Since the type-to-search box was not mapped to the Health Canada Licensed Natural Health Products Database (LNHPD), NHP were entered as free-text/numeric inputs. The interviewer also recorded information about the dosage, frequency, duration, start date, and indications for use. Since dose and frequency data were gathered, strength was not collected.
Figure 1. Input mapping by algorithm and manual processes in CLSA’s baseline Comprehensive cohort participants.
Note: Algo = algorithm; DPD = Drug Product Database; NHP = natural health product.
Drugs authorized for sale by Health Canada are listed in the Health Canada DPD (Health Canada, n.d.a), which contains information notably on product name, list of active ingredients, DIN, and World Health Organization (WHO) anatomical therapeutic chemical (ATC) classification. NHP licensed by Health Canada are listed in the Health Canada LNHPD (Health Canada, n.d.b), which contains information notably on product name, product’s medicinal ingredients, product’s non-medicinal ingredients, and natural product number (NPN). The NHP database does not include ATC codes. Both databases are updated nightly.
Algorithm recoding
In a second step, sequential algorithms were applied to map free-text (drug or NHP names) or numeric (DINs or NPNs) inputs to the products of the Health Canada drug and NHP databases (Figure 1). Seven algorithms were developed in a software algorithm approach independent of the sample data (Table 1). The algorithms were run sequentially such that once an input was matched, it was no longer considered in the remaining algorithms. For a given input, the first algorithm attempted to map the input to the drug followed by the NHP database before moving on to the next algorithm. The Direct and Code algorithms were run first since they only ever matched a single input to a single drug or NHP, while the Word and Simple algorithms at times found multiple matches. In cases of multiple matches due to numerous dosage strengths, the input was matched to the suitable drug or NHP with the lowest DIN or NPN.
Table 1. Developed algorithms
Note: DIN = drug identification number; NHP = Natural Health Product; NPN = natural product number.
Work was conducted using SQL (database scripting language) and PHP (general programming language). The Health Canada databases and CLSA data were loaded into a secure MySQL database using SQL. Some pre-processing was conducted on these databases before using PHP to enhance performance, increase speed of matching, and make the computer algorithms more efficient. For instance, the Simple algorithm compared the unmapped inputs to drug and NHP names from the Health Canada databases by ignoring non-alpha-numeric characters. This was done by removing the non-alpha-numeric characters from both the unmapped inputs and the Health Canada databases names, then comparing the two. It would be slow to transform the drug names in this way every time a comparison is made. Instead, all drug names were electronically converted during this pre-process step once and used by the algorithm every time a match was searched for. Another example is a list that was made of all identical drug and NHP names. The final version of the algorithm sequence and variables from the Health Canada databases are presented in the Supplementary Material.
As part of an iterative algorithm improvement approach, two pharmacists (L.D. and B.C.) independently recoded 40 unmapped drug and NHP inputs. The pharmacist-recoded inputs were compared to algorithm-recoded inputs during meetings of the research team, leading to algorithm refinement. This process of review – discussion – algorithm refinement was conducted three times for a total of 120 inputs, leading to two new algorithms: Predefined and No-units (Table 1). The greater complexity of recoding NHP inputs compared to drug inputs was identified early in this process and discussed throughout our work.
Manual recoding
In a third recoding step, following the application of the algorithms to the unmapped drug and NHP data, the remaining unmapped de-identified data were exported directly from the CLSA’s database to an Excel file for manual recoding by three pharmacy technicians (Figure 1). The same group of recoders conducted the recoding and validation work. The recoders’ work was supported by a set of decision rules (Supplementary Material) to assign selected NPNs for the most prevalent NHP inputs (e.g., NPN = 80083109 for calcium).
Spelling dictionary
As inputs were manually recoded, common misspellings were compiled into a dictionary and applied to future iterations of the computer algorithms. In the pre-processing stage, all inputs containing any of the misspelled words in the dictionary were replaced with the correct spelling before the algorithms were run (Figure 1).
Validation process
A validation sample of 100 Comprehensive cohort participants was randomly selected to evaluate the performance of the recoding algorithms and manual recoding. This sample included 352 free-text drug and NHP inputs for which a gold-standard recoded input was determined independently by two recoders with resolution of discrepancies by a pharmacist. A gold-standard recoded input could not be established for some inputs due to insufficient input information. Differing commercial products of the same generic drug or NHP were considered to be an agreement. After this first validation, the algorithms were further refined and validated in a second sample of 544 Comprehensive cohort participants with 1,407 unmapped free-text drug and NHP inputs. In this second validation, the gold-standard recoded input was established by a single recoder based on the measured recoders consensus in the first validation.
Analysis
Manual recoding was considered the gold standard for free-text inputs. The proportion of algorithm-correctly recoded inputs was calculated as the number of algorithm-correctly recoded inputs, based on the gold standard, divided by the number of algorithm-recoded inputs. In the primary analysis, the denominator included only the inputs for which a gold standard could be established in order to distinguish between drug and NHP. In a sensitivity analysis, the denominator included all algorithm-recoded inputs, regardless of gold-standard coding, for a more conservative estimate that cannot differentiate between drug and NHP.
Ethics approval
The CLSA was approved by the Hamilton Integrated Research Ethics Board (approval number 10-423, for the Comprehensive cohort) at McMaster University and the research ethics boards of all collaborating institutions.
Results
Mapping and recoding of drug and NHP inputs
Among CLSA’s 30,097 baseline Comprehensive cohort participants, 26,000 (86.4%) were using a drug or an NHP. Among drug or NHP users, a mean of 5.3 (SD 3.8) inputs per participant were documented for a total of 137,366 inputs. In the first of a three-step process, interviewers mapped 70,177 (51.1%) of the 137,366 inputs to a drug in the Health Canada DPD (Figure 1). Of the remaining 67,189 unmapped inputs (Figure 1), 3,247 (4.8%) were pre-processed by the spelling dictionary. In step 2, the Direct and Code algorithms recoded 10,657 (7.8%) drug and 10,072 (7.3%) NHP inputs. In step 3 (manual recoding), 10,185 (7.4%) drug and 33,923 (24.7%) NHP inputs out of the 46,460 (32.1%) remaining unmapped inputs were manually recoded (Figure 1). Insufficient input information resulted in an inability to code for 2,352 (1.7%) inputs (e.g., study drug and hypertension medication), made available to researchers as entered (Figure 1).
Algorithm and manual recoding validation
First validation sample
From the first validation sample, 352 free-text inputs were submitted to algorithm recoding and reviewed by two recoders (and pharmacist for non-consensus inputs) to establish a gold-standard recoded input. Of these 352 inputs, 12 free-text inputs were not recoded by the recoder nor the algorithms because of insufficient information. Of the remaining 340 inputs, 307 were recoded by the algorithms (Table 2). The Direct algorithm recoded the most (49.5%) inputs followed by the Word algorithm (22.5%). In the main analysis of the inputs for which a gold standard could be established, the Direct and Word algorithms correctly classified 97.9 per cent and 59.3 per cent of drugs and 96.2 per cent and 30.6 per cent of NHP inputs, respectively. In the sensitivity analysis of all algorithm-recoded inputs, the Direct and Word algorithms correctly classified 95.4 per cent and 39.1 per cent of inputs.
Table 2. Validation of algorithm recoding with manual recoding (gold standard) – first validation sample
Note: NHP = natural health product.
a Percent of the manually recoded drug inputs also recoded by the Direct algorithm out of the 147 manually recoded drug inputs.
b Percent of correctly recoded drug inputs by the Direct algorithm out of the 96 recoded drug inputs by the Direct algorithm.
c Percent of correctly recoded NHP inputs by the Direct algorithm out of the 53 recoded NHP inputs by the Direct algorithm.
d Percent of correctly recoded inputs by the Direct algorithm out of the 152 recoded inputs by the Direct algorithm.
Of the 352 drug and NHP inputs, consensus was reached by both recoders for 294 (83.5%) inputs. Of these 352 inputs, the recoders agreed that there was insufficient information to recode 21 inputs, excluded from the following subgroup analysis. Of the remaining 329 inputs, consensus was reached by the recoders for 156 (89.7%) of the 174 drug inputs and for 116 (74.8%) of the 155 NHP inputs. Based on these results, the second algorithms’ validation was conducted with a gold standard established by a single recoder. The recoders’ consensus was similar for algorithm-recoded inputs (83.4%) and non-algorithm-recoded inputs (84.4%).
Second validation sample
Of the 1,407 free-text inputs of the second validation sample, 27 were not recoded by the recoder nor the algorithms because of insufficient information. Of the remaining 1,380 inputs, 1,280 were recoded by the algorithms (Table 3). The Predefined algorithm recoded the most (44.8%) inputs followed by the Direct algorithm (29.0%). Modifications to the predefined algorithm for the coding of vitamins explains the increase in recoded inputs from the first to the second validation sample. In the main analysis of the inputs for which a gold standard could be established, the Direct and Pre-defined algorithms correctly classified 99.4 per cent and 86.4 per cent of drugs and 99.5 per cent and 78.2 per cent of NHP inputs, respectively. In the sensitivity analysis of all algorithm-recoded inputs, the Direct and Pre-defined algorithms correctly classified 94.6 per cent and 77.0 per cent of inputs. Following the second validation, the Code and Direct algorithms were selected for step-2 algorithm recoding of the unmapped free-text inputs of the baseline Comprehensive cohort participants.
Table 3. Validation of algorithm recoding with manual recoding (gold standard) – second validation sample
Note: NHP = natural health product.
a Percent of the manually recoded drug inputs also recoded by the Direct algorithm out of the 349 manually recoded drug inputs.
b Percent of correctly recoded drug inputs by the Direct algorithm out of the 171 recoded drug inputs by the Direct algorithm.
c Percent of correctly recoded NHP inputs by the Direct algorithm out of the 182 recoded NHP inputs by the Direct algorithm.
d Percent of correctly recoded inputs by the Direct algorithm out of the 371 recoded inputs by the Direct algorithm.
Discussion
We described a three-step process for the mapping of drug and NHP data to Health Canada databases that included algorithm recoding of 15.1 per cent of all drug and NHP inputs with high confirmation against gold-standard manual recoding. The developed algorithms have and will continue to save significant manual recoding time considering the large volume of CLSA drug and NHP data collected every 3 years over 20 years. The three-step process will enable the medications data collected from CLSA participants to be curated more efficiently and released as part of the CLSA research data platform for use by researchers. The process has the potential to be tested and applied with other large studies.
In the first of the three-step process, CLSA in person interviewers mapped 51 per cent of 137,366 drug and NHP inputs to the Health Canada DPD. In CLSA, the mapping of all drugs and NHPs, a much more extensive and diverse data set, contrasts from the mapping to a selection of 2,025 common medications in the multi-national, ASPREE clinical trial in older adults (Lockery et al., Reference Lockery, Rigby, Collyer, Stewart, Woods, McNeil, Reid and Ernst2019) and to a list of the 32 most common medications used in the Australian population in the 45 and Up study (Gnjidic et al., Reference Gnjidic, Pearson, Hilmer, Basilakis, Schaffer, Blyth and Banks2015).
In the second mapping step, two (Code and Direct) of the seven developed algorithms were selected for algorithm recoding of unmapped drug and NHP inputs. The limited number of selected algorithms highlights the need for a validation process to identify the challenging inputs in a specific data set. In our final validation sample, the Direct algorithm correctly classified 99.4 per cent of drug and 99.5 per cent of NHP inputs among the inputs for which a gold standard could be established. Similar validations of drug mapping/recoding have been reported by other groups. In the 45 and Up study, the automated coding of drug terms first to generic names using the Systematized Nomenclature of Medicine – Clinical Terms followed by coding to the WHO – ATC classification achieved positive predictive values above 95 per cent and sensitivity of 79 per cent at the exact ATC level with higher sensitivity values for drugs than vitamins and supplements (Gnjidic et al., Reference Gnjidic, Pearson, Hilmer, Basilakis, Schaffer, Blyth and Banks2015). The cleaning of drug names in the Food and Drug administration Adverse Event Reporting System (FAERS) database resulted in standardization of 95 per cent of drug name (Veronin et al., Reference Veronin, Schumaker, Dixit, Dhake and Ogwo2020). In another study on the FAERS database, drug name coverage of 93 per cent was achieved in the mapping to RxNorm standard code ingredients (Banda et al., Reference Banda, Evans, Vanguri, Tatonetti, Ryan and Shah2016). With highly structured inpatient pharmacy data from the GEMINI database from seven Canadian hospitals over 8 years, the use of existing Rx-Norm functionality resulted in sensitivity greater than 98.5 per cent and an F-Measure above 90.0 per cent in the standardization of 13 selected drug classes (Waters et al., Reference Waters, Malecki, Lail, Mak, Saha, Jung, Imrit, Razak and Verma2023).
In the third mapping step, 33.8 per cent of the remaining unmapped inputs were manually recoded with higher consensus for drug than NHP inputs. The mapping of the NHP inputs to Health Canada’s LNHPD adequately documents the product name as recommended by the CONSORT statement on herbal interventions (Gagnier et al., Reference Gagnier, Boon, Rochon, Moher, Barnes and Bombardier2006). It allows researchers using CLSA data to further detail the physical characteristics of the NHP such as the part of the plant used to produce the extract and the type of product used (e.g., fresh or dry) as suggested by the CONSORT statement. General NHP designations (e.g., multivitamins) were coded as per our decision rules.
Strengths and limitations
The main strength of our approach is the mapping/recoding of drug and NHP data to standardized information of Health Canada’s Drug and NHP Databases. The availability of these regularly updated databases was essential to this project. This linkage included the WHO ATC categories for drugs, a derived variable particularly useful for researchers using CLSA data. Our sequential approach limited the manual recoding to 33.8 per cent drug and NHP inputs. The main limitation of our approach is in the initial free-text entry of all NHP inputs and the 74.8 per cent consensus during manual recoding. Also, our approach would need to be adapted for drug and NHP data collection in other countries because of varying names.
Making the CLSA drug and NHP data available to researchers
CLSA data are currently available to approved public sector researchers in Canada and elsewhere. The data application process is described on CLSA’s website (http://www.clsa-elcv.ca), which also hosts the medication and NHP data support document providing a brief overview.
Ongoing developments
We continue to refine our collection and curation processes for medications data in the CLSA by exploring the linkage of the type-to-search box to Health Canada’s LNHPD for the mapping of NHP information by CLSA interviewers. The multiple brand name extensions generating an important number of options that could increase interviewers’ data collection time is a concern for NHP mapping. We are pursuing the refinement of the algorithms using new classification approaches and evaluating the integration of these refined algorithms to the type-to-search box to generate a list of possible matches to Health Canada’s LNHPD and limit the need for manual recoding.
Conclusion
We created an efficient three-step sequential process for drug and NHP data collection and curation in a longitudinal cohort as shown by the mapping of half of the drug and NHP inputs by the interviewers and algorithm recoding of 15.1 per cent of inputs. The accuracy of our approach was shown by the confirmation of algorithm coding compared to gold-standard manual recoding and recoders consensus for drug for the manual recoding process. Our approach has the potential to be applied by researchers using other large data sets requiring cleaning. We are pursuing the development of our approach for the data collection and mapping of NHP data to Health Canada’s LNHPD and integrating the algorithms into the day-to-day working of the next set of follow-up data collection periods in the CLSA.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0714980823000806.
Acknowledgements
The authors would like to thank the participants who give their time to the Canadian Longitudinal Study on Aging. The authors would also like to thank Helga Weigelin and Bevonie Brown, who conducted the manual recoding, Jean-Philippe Turcotte and Claudie Rodrigue for the data analysis, and Joanne Ho, Carol Bassim, and Kasia Makara for their support in coordinating this work.
Financial support
Funding for the Canadian Longitudinal Study on Aging (CLSA) is provided by the Government of Canada through the Canadian Institutes of Health Research (CIHR) under grant reference LSA94473 and the Canada Foundation for Innovation, as well as the following provinces, Newfoundland, Nova Scotia, Quebec, Ontario, Manitoba, Alberta, and British Columbia. The funders had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The CLSA is led by Drs Parminder Raina, Christina Wolfson, and Susan Kirkland. B.C. is a Junior 1 Research Scholar from the Fonds de recherche du Québec – Santé. L.G. is supported by the McLaughlin Foundation Professorship in Population and Public Health. P.R. holds the Raymond and Margaret Labarge Chair in Optimal Aging and Knowledge Application for Optimal Aging, is the Director of the McMaster Institute for Research on Aging and the Labarge Centre for Mobility in Aging, and holds a Tier 1 Canada Research Chair in Geroscience.
