This HIV case-based surveillance pilot was conducted between July 2015 and December 2015 in 124 facilities in Kisumu and Siaya counties. The facilities were selected to represent a variety of settings such as levels of care (dispensary, health center, subcounty, and county referral), use of an EMR versus paper records, and size of the patient population. Data were collected retrospectively to allow for at least four months of follow-up time from initial diagnosis, entry into care, or ART initiation within the study period. Data were collected by subcounty AIDS and sexually transmitted infections (STI) coordinators and Kenya Medical Research Institute (KEMRI) surveillance officers, and in some cases, facility staff. Data were entered from paper medical records and registers into the customized data entry platform for cases newly diagnosed or newly enrolled in HIV care from January through June 2015 using Android-based tablets and a standardized HIV case report form. Surveillance officers were trained in data collection using tablets and provided with login credentials. All surveillance officers signed a data confidentiality statement. As collected data contained patient names and other patient identifiers they were encrypted before transmission via a dedicated virtual private network in real-time to a server hosted on the Amazon cloud computing service. The staff at the National AIDS and STI Control Program (NASCOP) managed the data [19].

A case was defined minimally to include the date of diagnosis, age at diagnosis, gender, first name, and surname. Cases originated from the following 3 scenarios and analytical frameworks relating to the HIV care cascade. The first scenario is within HIV testing services (HTS). This scenario accommodates cases found within the same facility (cases that were tested at the facility and retested at the same facility hence having different dates of diagnosis). It also included cases that moved to a different facility (cases that tested at one facility and retested at a different facility). The second scenario is HTS-care. This accommodated HTS-to-care scenario in which cases were tested and linked within the same facility. It also included HTS-to-care cases that would be tested at one facility and then linked to care in a different facility. These 2 scenarios accounted for movement of persons diagnosed with HIV and accessing care within the same facility and clients that may test at one facility and access care in a different facility. The third scenario is within care scenarios. This included referrals and linkages from one facility to another. Similar to HTS-to-care linkage scenarios some cases had enrolled into care in one facility and throughout care transferred to another facility. However, HTS was not a source of data for the diagnosis information, and hence we did not have any testing location information for these cases.

Methods for data collection are described in the HIV case-based surveillance pilot report [19]. Briefly, data were extracted prospectively for everyone newly diagnosed or enrolled in care in a given 6-month period in the participating facilities and subsequent updating of sentinel events for those individuals. At the end of the pilot, we had 12,260 records but excluded 100 which had a missing date of diagnosis and 3 which had a missing date of birth before matching (Figure 1).

We created analytical groups—also called “blocking” according to the scenarios described in the study setting before carrying out matching analyses to allow for comparability and faster processing,

Before carrying out matching processes, we standardized patient identifying fields used in matching. First, all blank spaces, commas, apostrophes, and dashes were stripped from first names middle names and surnames. Second, all string fields were converted to lower case. A Soundex [20] was created for the first names in all records since the first names are mostly of English origin. Third, we created double metaphone for middle names and surnames. Fourth, the year of birth was standardized to a four-digit number.

A potential patient identifier for the deduplication process is CCC number which is a unique patient number assigned at first clinical encounter once an HIV-infected patient has gone through triage and is ready for enrolment into a facility-managed HIV program. The CCC number is an 11-character code comprising a 5-digit unique facility code followed by a separator and a 5-digit sequentially facility-assigned unique number. We standardized CCC numbers to consider variations in recording (eg, use of spaces, slashes, dashes, adding leading zeros, and commas).

We used the following fields in deterministic matching (1) the first name, (2) surname, (3) gender, and (4) year of birth. To reduce mismatching due to variation in spellings of English first names, we used Soundex. We then created a “pseudo-unique key” combining the resulting Soundex values as well as gender, surname, and year of birth. The CCC numbers were used to match care records that were missed by using the “pseudo-unique key.”

We separated the data according to the “blocking” scenarios described in the deterministic process. These blocking scenarios are necessary so that comparisons are made among potentially related records. We used R (an open-source software) in our study since it provided programming flexibility to implement the matching string preparation and matching process. We created a matching key field by including the data elements (1) first character of gender at birth, (2) Soundex of the first name, (3) secondary double metaphone of middle name, (4) secondary double metaphone of surname, and (5) year of birth. This produced strings such as “fF465aknannk1983,” “fI650aknkannk1990” (where middle name secondary double metaphone was available), and “fG620ans1994” (where secondary double metaphone of the middle name was unavailable). We then implemented Jaro and Jaro-Winkler string matching and Levenshtein and Damerau-Levenshtein string edit distance algorithms in the R stringdist package [21,22]. String score-based matching was conducted using ratios of matching strings, and a penalty was applied for the first 4 characters when Jaro-Winkler algorithm is used as in the formula (Figure 2) where d_j is the Jaro-Winkler distance score, m is the number of matching characters, |s₁| is length of string 1, |s₂| is length of string 2 and t is half the total transpositions or the number of matching (but different sequence order) characters divided by 2. String edit distance calculations, on the other hand, quantify how different 2 strings or words are to one another by counting the minimum number of deletions, insertions, substitutions and transposition operations required to transform 1 string into the other. Score-based methodologies are based on the Fellegi-Sunter linkage rule that classifies a record pair as matching or nonmatching [11]. The score level to determine a match is determined a priori or based on experience by the user and dependent on the setting. For our case, a score of 98% and above was considered sufficient to determine a match. When we implemented the Jaro and Jaro-Winkler methods, we set a standard penalty factor to 0.1. This penalizes matches based on similarity at the beginning of the string to give favorable ratings to strings that match from the beginning for a set prefix length of up to 4 characters according to Winkler and Cohen [11,23]. The penalty factor is added to discount matches that are found based on up to a maximum of first 4 characters since in string writing, the person recording is more likely to make an error after the first 4 characters. We considered the 4 weights (ω) applicable to the Levenshtein and Damerau-Levenshtein methods (1) deletion (ω=0.8), (2) insertion (ω=0.8), (3) substitutions (ω=1), and (4) transposition (ω=0.5). For the Levenshtein method, the penalty for substitution is ignored [22].

Due to possibilities of age variations for the same person accessing HTS and care services at differing periods, the numeric comparator age, with a variation of not more than 12 months within identified matches was considered sufficiently close for confirming a match. We compared deterministic and score-based processes for unique case identification regarding the number of matches yielded and the deduplication extent achieved within the scenarios. We also assessed match yield when HTS and HTS-care records were treated as mutually exclusive versus as a combined set. Regardless of approach, total yield was a sum of duplicates from all scenarios.

Based on the date of HIV diagnosis, we carried out extra steps to determine how to retain unique records after the matching process. If the retained and duplicate records had conflicting dates of diagnosis, we retained the records with the earliest date of diagnosis. For retained records, we maximized completeness of data for all fields by comparing with the duplicate records. Whenever a retained record had missing data that was in duplicate record, an append merge was carried out to overwrite missing values with the nonmissing value from the matched record.

Ethical approval was obtained from the KEMRI (SSC #2827) and the Office of the Associate Director for Science, Centers for Disease Control and Prevention (CDC) with tracking #2014-136. Access to data used in these analyses was password protected, and all study coordinators, data abstractors, and analysts signed a confidentiality form.

A total of 12,260 records were collected. We excluded 100 (0.8%) records due to missing dates of diagnosis, and 3 (0%) missing the date of birth (Figure 1). The final data set used for the matching exercise included 12,157 records representing adult and pediatric cases. From these records and before data deduplication, 33.5% (4073/12,157), 9.0% (1091/12,157), and 57.5% (6,993/12,157) corresponded to HTS, HTS-care and within care scenarios respectively. In Table 1, completeness and uniqueness of variables used to construct score-based matching string are presented. In the entire data set, gender, year of birth, first name and surname were 100% complete while the middle name was missing for 38% of the records. First names were less unique than surnames: 8.2% (1002/12,157) versus 19.1% (2321/12,157). When Soundex was applied to standardize the English first names, 273/12,157 (2.2%) remained unique compared to 1002/12,157 (8.2%) of the original unstandardized format. When secondary double metaphone was applied to standardize the middle and surnames, 2.6% (316/8772) and 3.1% (373/12,157) respectively remained unique compared to 13.1% (1150/8,772) and 19.1% (2321/12,157) of the original unstandardized format. The similarity of names varied by setting (Table 1).

Out of the 12,260 records, 12,157 (99.2%) were used in the analyses. Using the deterministic method, 67/12,157 (1.6%) records were matches in HTS, 164/12,157 (15.0%) in HTS-care, and 204/12,157 (2.9%) in the care-only scenario. This yielded a total of 435/12,157 (3.6%) matches and 11,722 unique cases across the testing and, care and treatment cascade (Table 2).

Overall, of the score-based methods, Jaro-Winkler yielded the most duplicate records (686/12,157, 5.6%), Jaro yielded the fewest (546/12,157, 4.5%), and both Levenshtein and Damerau-Levenshtein yielded the same number (563/12,157, 4.6%). Specifically, duplicate records yielded by method were: (1) Jaro 5.7% (234/4073) within HTS, 0.4% (4/1,091) in HTS-care, and 4.4% (308/6993) within care, (2) Jaro-Winkler 7.4% (302/4073) within HTS, 0.5% (6/1091) in HTS-care, and 5.4% (378/6993) within care, (3) Levenshtein 6.4% (262/4073) within HTS, 0.4% (4/1091) in HTS-care, 4.2% (297/6993) within care, and (4) Damerau-Levenshtein 6.4% (262/4073) within HTS, 0.4% (4/1091) in HTS-care, and 4.2% (297/6993) within care.

A comparison of Jaro-Winkler yield for mutually exclusive and data sets that were combined across the scenarios is presented in Figure 3. When scenarios were treated as mutually exclusive, Jaro-Winkler score-based matching algorithm yielded 7.0% (302/4073) matches in the HTS scenario, 1% (6/4073) in the HTS-care scenario compared to a higher yield of 7.1 % (368/5164) when the 2 scenarios were treated as 1 block.

Based on the outcomes of the score-based matching process, we propose a procedure comprising of 7-steps that is easy to apply to quickly match and link unique cases across HIV care settings (Textbox 1). To decide whether or not to use demographic data matching, we propose a decision model (Figure 4).

Universal health care identifiers are recommended and ideal for patient-centered monitoring and CBS [24,25]. However, in low resource settings, their use is limited. In the interim, demographic data score-based matching algorithms can play an important role in improving the quality of CBS data as well as patient-centered care. We have demonstrated that score-based methods succeeded in patients matching and identifying more matches compared to the deterministic process. It is possible to match cases, merge sentinel events, and enhance the completeness of individual deduplicated data using this process. Consequently, this improves accuracy in CBS and other longitudinal encounter data. The process also has a dual utility of allowing better care coordination and patient management at the facility level and improved HIV surveillance at a higher subnational or national level. The matching process can be inbuilt in EMRs and at patient registries to allow for lookup of already registered patients at the facility level. This may improve processes, patient flow and avoid unnecessary double entry. We also demonstrate that we can do enough matching in the absence of a UHID to move ahead with CBS implementation in low-resource settings such as Kenya. As such, lack of a UHID should not stifle movement towards the use of CBS.

Our study compared 4 variants of score-based string-distance matching methods. The Jaro-Winkler distance method was found to perform better in score-based matching since it gave the best yield while considering common spelling mistakes and logical combination of demographic fields. In developed countries, it has been shown that about 5% to 10% of medical records may be duplicate [26], which compares well with our results. Jaro-Winkler has been proposed as a method over other string-matching algorithms since it was designed with relatively short strings in mind [21], hence may be best suited to our setting. In addition, it works well when the name beginnings are the same [27]. For that reason, we standardized beginning of the match strings by using a Soundex of the English names and using secondary double metaphone of middle and surnames. Further, a decision was made to add the first character of gender at birth to the beginning of the string to improve the accuracy of the matching score.

Although we used R in our analyses and matching process, open source software such as CDC Registry Plus Linkplus [28], which was originally developed for cancer registries has been explored in low-resource HIV care settings for example in Haiti [29]. Other Web-based applications that have utility for fuzzy matching and record cleaning, for example, Freely Extensible Biomedical Record Linkage [6], may have potential. However, post-match processing is necessary to achieve a high degree of true matches. A certain degree of human adjudication may be necessary especially when testing the algorithms. Users of off-the-shelf solutions such as Linkplus should take caution since many mismatches may be likely to be true matches [29]. The use of current English name-based Soundex algorithms is not appropriate for Kenyan names. In creating unique identifiers that contain a Soundex component, variations of the first name can yield a different Soundex since the first character is always part of the Soundex [20]. A visual inspection of matches based solely on Soundex of first and surname showed a high false-positive rate. Research on how to construct a Soundex algorithm for Kenyan names may be useful as has been successfully done in Japan, India, and South Africa [30-32]. We determined that using a double metaphone had discriminatory power for Kenyan names and hence we used it for middle and surnames.

Our study has several limitations. First, the choice of a combination of several fields for a concatenated “pseudo-unique key” may not be optimal. However, we developed the matching string taking advantage of existing identifiers in our data. First names in Kenya are usually English baptismal or anglicized names. We took advantage of this to standardize names that are misspelled using Soundex. Other challenges include manual transcription errors during patient transfers and assigning of new numbers for transfer-in patients. Despite these limitations, we were able to merge the cases based on the names, gender, date of birth and CCC number in the within care scenario and hence identify potential matches in the deterministic process. Finally, many studies have applied common measures of validity such as positive predictive value, sensitivity, and specificity [33]. Unlike those studies, we did not have a gold standard for comparison in the pilot.

The choice of which string distance score-based algorithm to use largely depends on the nature of the match strings and the nature of typographic errors [21]. Choice of the matching string is therefore important. For example, deterministic matching yielded more duplicates for the HTS to care scenario (15%) compared to 4.6% to 7.1% across the score-based methods. This may be because a rigorous manual assessment of possible matches was done using the CCC numbers such that matches within the HTS to care scenario were more efficiently captured. Minimalistic demographic fields were used in score-based matching across all scenarios, and the CCC number was not included in the process.

There has been an ongoing discussion and suggested approaches for countries to consider in developing UHIDs [17,34]. If, and when implemented, UHIDs would have the highest potential to mitigate challenges with a unique identification and record linkage for an expanded national CBS system. This benefit extends to other health sectors as countries move towards universal health care. The recent World Health Organization guidelines for patient-centered monitoring advocate for using unique patient identifiers instead of names [25]. However, where there is no UHID, a unique patients’ deduplication algorithm based on available demographic data is necessary and feasible. Such an algorithm would improve monitoring of the HIV epidemic including the UNAIDS Fast-Track 90-90-90 targets.

We propose a stepwise process that builds up from first identifying data sources and blocking scenarios. This should be followed by an examination of the data quality using completeness as a measure coupled with quality improvement measures through routine data quality audits. The next step involves developing a matching key, lower-level deduplication and finally cross-examination, validation and sending of CBS data to the national level for surveillance. Although validation of the score-based approach is a necessary extra step, this may be best done with data sets from settings where a gold standard is available such as those utilizing biometric finger vein technologies for patient identification. Given that these settings are rare, we suggest that programs identify a percentage that best suits their setting and resources for validation purposes. A decision model such as the one presented in Figure 4 may help programs to decide whether or not to use demographic data matching. Comparing score-based matches to gold standard data in Kenya and similar settings offer an opportunity for future work in search of alternatives for patient matching. In the meantime, score-based demographic data matching has utility for improving the quality of data in monitoring the 90-90-90 cascade and in other health care settings where patients are longitudinally followed.