Digital Healthcare Empowering Europeans R. Cornet et al. (Eds.) 2015 European Federation for Medical Informatics (EFMI). This article is published online with Open Access by IOS Press and distributed under the terms of the Creative Commons Attribution Non-Commercial License. doi:10.3233/978-1-61499-512-8-135 Design, development and first validation of a transcoding system from ICD-9-CM to ICD-10 in the IT.DRG Italian project Vincenzo DELLA MEA a,1, Omar VUATTOLO a, Lucilla FRATTURA b, Flavia MUNARI a, Eleonora VERDINI c, Loris ZANIER d, Laura ARCANGELI e, Flavia CARLE e a Department of Mathematics and Computer Science, University of Udine b Central Health Directorate, Classification Area, Friuli Venezia Giulia Region, and Italian WHO-FIC Collaborating Center, Italy c Health Information System Service, Emilia Romagna Region, Italy d Central Health Directorate, Health Information System Service, Friuli Venezia Giulia Region, Italy e Ministry of Health, VI Office, Rome, Italy 135 Introduction Abstract. In Italy, ICD-9-CM is currently used for coding health conditions at hospital discharge, but ICD-10 is being introduced thanks to the IT-DRG Project. In this project, one needed component is a set of transcoding rules and associated tools for easing coders work in the transition. The present paper illustrates design and development of those transcoding rules, and their preliminary testing on a subset of Italian hospital discharge data. Keywords. International Classification of Diseases, Clinical Coding, Health Information Systems In Italy, ICD-9-CM is currently used for coding health conditions at hospital discharge. In order to introduce ICD-10 in morbidity coding like in other countries (e.g., [1]), and revise the overall case mix classification system, since 2010 a national project has been funded led by the Italian Ministry of Health and Emilia Romagna Region ( IT-DRG Project ). It involves the Friuli Venezia Giulia Region (as Italian WHO Collaborating Center for the Family of International Classifications) to update ICD-10 and the Lombardia Region to update the interventions and procedures classification. In order to evaluate the impact of ICD-10 introduction in Italy, the Italian WHO- FIC collaborating center has been active on the translation of ICD updates, the update of ClaML files [2] and development of web tools and services. A preliminary version of the ICD-10 2013 Italian version was used as the reference version to transcode administrative hospital discharge data by a specifically developed web tool named 1 Corresponding Author.
136 V. Della Mea et al. / Design, Development and First Validation of a Transcoding System TransIT. This was developed to make the transition easier for coders that already know ICD-9-CM. 1. Methods 1.1. Transcoding rules design and development TransIT transcoding rules were obtained initially by processing the original ICD-9- CM 2007 to ICD-10-CM 2007 transcoding rules [3,4], complemented by rules identified by classification experts to take into account the differences between ICD-9- CM and ICD-10 (in particular, the dagger/asterisk convention) and also between ICD- 10 and ICD-10-CM, which contains much more entities. These rules classify ICD-9-CM codes as transcodable and not transcodable. Each rule in turn is classified on 2 axes: approximate or exact, and simple (an ICD-9-CM code corresponds to a single ICD-10 code) or (each ICD-9-CM code corresponds to a set of ICD-10 codes). Finally, each ICD-9-CM transcodable code may have single or many alternative choices. We identified a three-steps process: (i) automatic adaptation of ICD-10-CM rules to ICD-10; (ii) improvement of rules through lexical techniques and (iii) expert review and refinement. Since ICD-10-CM is an extension of ICD-10, at step (i) we checked automatically, for each ICD-9-CM code, each ICD-10-CM target code: if the target code was not a leaf node in the ICD-10 classification, it was replaced by its leaf descendants; if the target code was an extension of ICD-10, it was truncated to its closer ancestor available. At step (ii), we tried to improve transcoding rules applying lexical techniques, looking for exact text-matching between ICD-9-CM titles and ICD-10 titles and inclusions. When the match between an ICD-9-CM code is found on a ICD-10 leaf code, old transcoding rules were dropped and replaced by an exact simple one. If the match is found on a non-leaf ICD-10 code, old transcoding rules were dropped and replaced by a set of approximate simple transcoding rules. Finally (step (iii)) an expert coder reviewed a number of transcoding rules, identified automatically as critical. The aim was to verify the rules and eventually improve them, reducing the number of choices in a safe way for a better support of future coders. 1.2. Transcoding system A system has been designed and preliminarily developed to allow transcoding in three different ways: online, through a web-based interface; directly from coding software, through REST web services; offline, by means of transcoding tables in different formats (CSV, XML, JSON).
V. Della Mea et al. / Design, Development and First Validation of a Transcoding System 137 1.3. Validation The full set of Hospital discharge forms data for 3 Italian regions on 2011 and 2012 was obtained from the Ministry of Health to test the transcoding system. TransIT was then used to transcode Hospital Discharge Forms (SDO) ICD-9-CM codes from the available database. Each SDO record contains one code for the main condition and up to 4 codes for other conditions. Data on the kind of rules used for transcoding were collected, in particular for distinguishing single choice from multiple choice, being the latter the most time consuming option for coders. Table 1 shows details on the sample. Table 1. Sample size of the Italian Hospital Discharge Forms considered for validation Region Year Total records Main conditions Other conditions Friuli Venezia Giulia 2011 197,664 197,663 235,566 2012 198,225 198,224 240,074 Veneto 2011 729,748 728,210 2012 697,020 696,991 596,233 587,969 Emilia Romagna 2011 787,142 787,142 2012 764,511 764,511 1,023,630 1,006,775 Total 3,374,323 3,372,741 3,690,247 2. Results 2.1. Transcoding rules and system The distribution of transcoding rules for each step is shown in Table 2, according to the steps described in section 1.1. Table 2. Number and distribution of transcoding rules from ICD-9-CM to ICD-10 Step Exact Exact simple simple Exact simple (i) 2372 33 5 7,622 1,461 442 340 (ii) 2521 43 16 7,438 1,494 442 329 (iii) 2533 93 17 7,537 1,355 541 182 The result of the final step is a set of 12,258 rules, of which a total of 10,704 (87.3%) is single choice, either exact, approximate or. However, these numbers have to be read in the context of real world data, to fully understand how much these rules are used. For this aim, the validation phase has been designed. The system embedding the rules has been implemented using PHP5, MySQL, JQuery, Twitter Bootstrap and successfully tested from a technical point of view. 2.2. Validation A total of 3,374,323 SDOs were analysed, of which 3,372,741 contained a main condition, with a total of 3,690,247 secondary conditions. The number of different ICD-9-CM codes used was 10,987 (88.4% on a total of 12,435). Table 3 shows details
138 V. Della Mea et al. / Design, Development and First Validation of a Transcoding System on the transcoding rules used, grouped according to the most relevant features (exact, approximate,, and multiple choices). Looking at both main and secondary conditions, a large number of SDOs ICD-9- CM codes (86.36%) was transcoded automatically, that is, transcoding provided just one option, either exact (35.86%), approximate (50.18%) or (0.32%). The remaining 12.86% of SDO ICD-9-CM codes needed manual intervention, since transcoding provided more than one option. 401 codes could not be transcoded (0.006%). When analyzing details of codes that could not be transcoded, a number of coding mistakes were found (mostly: intermediate level categories and groups that cannot be used for SDOs coding conditions according to coding rules, but yet having been used by coders). Table 3. Transcoded Hospital Discharge Forms by Region,Year, and type of transcoding rule Region Year Exact simple simple Composite Multiple not (1 choice) (1 choice) (1 choice) choices transcoded Friuli V. Giulia 2011 161,524 212,018 1,951 54,308 0 2012 163,131 215,795 1,971 53,911 0 Veneto 2011 473,204 659,351 4,613 177,388 2012 457,646 642,714 4,977 170,545 199 201 Emilia Romagna 2011 644,378 919,009 4,468 228,357 2012 632,884 895,487 4,812 223,810 1 0 Total (N) 2,532,767 3,544,374 22,792 908,319 401 Total (%) 35.86% 50.18% 0.32% 12.86% 0.006% Again to understand the amount of work needed by coders switching from ICD-9- CM to ICD-10, we identified the most used ICD-9-CM codes and the rules needed for their transcoding. In particular, we examined the 100 most used ICD-9-CM codes, which covered 42% of SDOs. Of these, 44 were coded exactly, 46 approximately, and 10 with multiple choices (Table 4 shows only the first 6 codes, accounting for about 10% of total codes). Codes not transcoded were coding mistakes (non-leaf codes). Table 4. The most used ICD-9-CM/ICD-10 categories in the SDO sample ICD-9-CM rule type ICD-10 Count 427.31 Atrial fibrillation Exact simple I48.0 160,051 401.1 Benign essential hypertension Appr. simple I10 154,454 V30.00 Single liveborn, born in hospital, delivered without Appr. simple Z38.0 119,093 mention of cesarean section 250.00 Diabetes mellitus without mention of complication, Appr. simple E11 99,086 type ii or unspecified type, not stated as uncontrolled 401.9 Unspecified essential hypertension Exact simple I10 72,246 V58.11 Encounter for antineoplastic chemotherapy Appr. simple Z51.1 67,694 Discussion The proposed transcoding system provides three different ways of using it, aimed at different environments. The web-based interface is suitable for direct use by coders, and it can be useful for training at first, but also when coding outside information systems, in the most traditional way. Transcoding tables and web services are meant for coding systems that could embed them in their interface. While transcoding tables, like
V. Della Mea et al. / Design, Development and First Validation of a Transcoding System 139 USA GEMs, represent the usual solution, web services might provide always updated transcoding rules. While the system has been developed with the coding scenario in mind, it will be also provided for the transcoding of historical series. In this case, no context knowledge is typically available for multiple choices, thus they should be reduced to single choice, by addressing coding towards a residual. Italy already switched from ICD-9 to ICD-10 for mortality coding when requested by WHO [5]. However, for morbidity coding ICD-9-CM is still used, together with ICD-9-PCS for coding procedures, both being a basis for DRG classification. This when part of the world already codes morbidity in ICD-10 [1,6], and WHO is already working on ICD-11. The IT-DRG Project is aimed at providing new classifications for all of the above mentioned components, i.e., diagnoses, procedures, and DRGs. For what regards diagnoses, the transition from ICD-9-CM to ICD-10, based on these preliminary data analysis, could be less difficult than supposed, because a large number of ICD-9-CM codes can be easily transcoded to one single ICD-10 code, leaving a manageable 13% of codes to be chosen by coders among a small set of options. However, training is needed for coders to understand the differences between the two ICD versions, in particular when involving the dagger/asterisk mechanism, which is not present in ICD-9-CM. The lack of the E codes in current SDO coding rules is another issue to address in ICD-10 V-W-X-Y codes implementation in Italy. Acknowledgements The present work has been done as part of the "Progetto It.DRG", funded by National Health Service 2004 to realize strategical objectives under the National Health Plan, and as part of the agreement between Italian Ministry of Health and Friuli Venezia Giulia Region, 2010-2012; 2013-2015 to support National Health Service in implementing WHO classifications. Work of VDM and OV at the Dept of Mathematics, University of Udine was funded by the Region Friuli Venezia Giulia, to support research activities on WHO classifications, 2014-2016. References [1] Kerry I., Peasley K., Roberts R. Ten Down Under: Implementing ICD-10 in Australia. Journal of AHIMA 2000, 71 (1): 52-56. [2] Frattura L., Grippo F., Frova L. The collaborative effort to implement updated classifications: the lesson learned in developing and using web tools and services to translate, update, browse, and publish ICD- 10. WHO-FIC Network Annual Meeting, Barcelona 2014 [3] Butler R. The ICD-10 General Equivalence Mappings. Bridging the translation gap from ICD-9. J AHIMA. 2007 Oct;78(9):84-5. [4] National Center for Health Statistics. 2007 release of ICD-10-CM, General Equivalence Mapping files. http://www.cdc.gov/nchs/data/icd/diagnosisgems_2007.zip [5] Brocco S, Vercellino P, Goldoni CA, Alba N, Gatti MG, Agostini D, Autelitano M, Califano A, Deriu F, Rigoni G, Cassinadri MT, Garrone E. [«Bridge Coding» ICD-9, ICD-10 and effects on mortality statistics]. Epidemiol Prev. 2010, 34(3):109-19. [6] Jetté N, Quan H, Hemmelgarn B, Drosler S, Maass C, Moskal L, Paoin W, Sundararajan V, Gao S, Jakob R, Ustün B, Ghali WA. The development, evolution, and modifications of ICD-10: challenges to the international comparability of morbidity data. Med Care.2010, 48(12):1105-10