SnoMAP: Pioneering the Path for Clinical Coding to Improve Patient - - PDF document

SnoMAP: Pioneering the Path for Clinical Coding to Improve Patient Care

Michael LAWLEYa, Donna TRURANa, David HANSENa, Norm GOODa,, Andrew Staibb, Clair Sullivanb

aAustralian eHealth Research Centre, CSIRO bPrincess Alexandra Hospital, Clinical Excellence Division Queensland Health

• Abstract. The increasing demand for healthcare and the static resources available

necessitate data driven improvements in healthcare at large scale. The SnoMAP tool was rapidly developed to provide an automated solution that transforms and maps clinician-entered data to provide data which is fit for both administrative and clinical purposes. Accuracy of data mapping was maintained.

Introduction

The healthcare system is undergoing rapid digital transformation. The initial primary driver for this digitisation

• f health care delivery is increased efficiency and quality at the point of patient care. However, increasingly

clinicians and system managers are seeing the potential for secondary use of the clinical data collected drive health system improvements. Increasing demand for healthcare in the face of static resources has reinforced this need for digital solutions enabling data driven decision making in healthcare. Australia has only recently delivered its first tertiary digital hospital with an integrated electronic medical record (EMR)[1]. During the rollout of this EMR however, it became clear that Australia has a clinical coding dilemma. The rich clinical data coded by the clinicians did not meet administrative coding requirements for government funding

• f the hospital. The decision had been made at a state level to use a clinically useful code set for the EMR

(Systematised Nomenclature of Medicine: Clinical Terms, Australian Extension (SNOMED CT-AU)) but the government required reporting of this data in a different, administratively useful, code set (International Statistical Classification of Diseases and Related Problems, Australian Modification (ICD-10-AM)). In order to retain funding for the new digital hospital, a strategy had to be developed to rapidly and accurately transform the clinically useful code set (SNOMED CT) into a different administratively appropriate code set (ICD-10-AM).The discrepancy between clinician-entered SNOMED codes and administratively required ICD codes for an inpatient stay could be dealt with by clinical coders manually entering ICD codes based on clinical information in the EMR. This was the planned mitigation strategy for situations such as inpatient hospital admissions where manual coding from the paper record already existed. However, not all hospital attendances were subject to manual entry of clinical codes by professional coders. There are over 1.5 million attendances to Emergency Departments per year in Queensland[2]. Clinical coding for these attendances was entered by clinicians at the point of care, and submitted directly to the central agencies as part of a minimum dataset for performance and financial [3-5]. Either an additional, manual coding step or a tool that allowed rapid, accurate mapping of the full range of SNOMED codes in use into ICD codes was required. Existing tools were inadequate due to lack of code coverage and accuracy, so we had to rapidly develop a solution. This aim of this paper is to: 1. Technical Brief:  define the current tension between clinically useful data sets and administrative data sets  provide a detailed description of the tool we delivered 2. Describe the implementation processes locally and across other sites

3. Outline the clinical care impact  describe a post hoc analysis showing funding levels and accuracy were maintained after transforming the dataset.

Technology Brief

Without a rapidly developed solution, the hospital could potentially face funding cuts. The timeline was short with the entire development occurring over four months. The brief was to create an automated solution which transforms and mapped clinician-entered data to provide data which was fit for both administrative and clinical purposes. The tension between datasets In Australia, the coding systems most widely used in health services are ICD-10-AM[6] and SNOMED CT- AU [6 7]. ICD-10-AM is a statistical classification designed for encoding information about inpatient episodes

• f care[7]. It is not suitable for use by clinicians, at the point of care or for clinical record documentation[8];

its categories often combine multiple conditions into a single code so that the specific condition is unknown. The statistical nature of ICD-10-AM encoded data best suits population health surveillance (morbidity and mortality)[7 9-11], and as a basis for financial modelling and resource utilisation[12 13]. ICD-10-AM is the mandated coding system for secondary data collections such as the Admitted Patient Care National Minimum DataSet [4] and the Non-Admitted Patient Emergency Department National Minimum Dataset[4]. SNOMED is an international standardised, multilingual vocabulary of clinical terminology that is designed for use by physicians and other health care providers to document patient medical records at the point of care in all healthcare settings. SNOMED CT includes synonyms and definitions with unique identifiers capable of being electronically exchanged between healthcare providers. SNOMED provides clinically relevant, very specific, and descriptive terms suited to care delivery. SNOMED encoded data collected at the point of care can be analysed, aggregated and re-used for multiple purposes [14 15], including safety and quality of care reviews and performance metrics. The Australian extension provides local variations and customisations of terms relevant to the Australian healthcare community. These two coding systems are different, intended for different user groups, and for different purposes. One does not replace the other but both must operate in harmony within the health information eco-system, where both clinical and statistical data must be accurate and meaningful and reflect the patient, clinician and health service experiences[16]. Existing Mapping Systems Mapping has long been regarded as the most effective way of allowing SNOMED to be implemented in EMRs while co-existing in a health information administrative environment that has traditionally used ICD-10-AM [15]. The maps previously developed for use in ED settings (2009) were inadequate, did not reflect the terminology preferences of clinician users, were not maintained and updated, and did not perform under Activity Based Funding protocols[17]. New comprehensive, version-controlled maps were required to allow PAH data to be mapped, audited and submitted in each monthly reporting cycle. Table 1 provides a description of the maps made available for this case study and Table 2 compares the previously released maps which cover only a subset of SNOMED concepts – those belonging to the Emergency Department Reference Set (EDRS) [15]. Table 1 AEHRC maps between SNOMED CT AU and ICD-10-AM 9th edition (2014) Reference Set SNOMED CT AU ICD-10-AM 9th edition 2014 Concepts 79, 639 10,018 Descriptions/synonyms 1,414,009 na

Table 2 Maps between Emergency Department Reference Set (EDRS), a subset of SNOMED CT AU, and ICD-10-AM 6th edition Emergency Department Reference Set ICD-10-AM 6th edition 2009 Concepts 5,760 3,055 Descriptions/synonyms na na Features that distinguish SNOMAP are: (i) the much broader scope of SNOMED CT content that maps to a greater number of ICD-10- AM codes, allowing clinicians to use the standard vocabulary however they need in order to properly document patient records; previously only an ‘approved’ subset was available; (ii) the inclusions of descriptions/synonyms permit point-of-care users to work with their favourite words at the interface, and links to concepts and hence to ICD-10-AM are managed for them in the background; data analysts do not have to convert from SNOMED CT-AU descriptions to concepts for mapping purposes; and (iii) the SNOMED CT-AU and ICD-10-AM releases and editions are current and are maintained. (iv) As can be seen in Tables 1 and 2, the maps themselves are mostly of the kind many-to-one, where the statistical classification (ICD-10-AM) gathers together in a coded category many more specific and descriptive SNOMED concepts. SnoMAP Solution The maps are deployed within a web-based tool called SnoMAP. Data managers upload patient data encoded in SNOMED. SnoMAP then processes that data into ICD-10-AM code. Any patient cases that do not achieve a map are flagged for follow up by data managers. This feedback loop supports data quality assurance at the local hospital level, before any erroneous data is submitted to the State repository. Because SnoMAP works very efficiently, data analysts can manage their data submissions in near real time.

Implementation Processes

The index hospital provided de-identified patient data collected over a three-year period. Each data extract included patient encounters between January and June for each of 2014, 2015, and 2016. This represents two quarterly statutory reporting time periods, covers about 5,000 patient cases per month, and any seasonal influences on patient health complaints are held constant. Data extracts from 2014 and 2015 were encoded originally in ICD-10-AM through the now retired Emergency Department Information System (EDIS). Data from 2016 was encoded directly in SNOMED CT-AU deployed through the newly introduced EMR system. These three data extracts enabled pre- and post-implementation comparisons of the influence of mapping on data outcomes under Activity Based Funding protocols. Validation Method Patient data for January to June 2014 and 2015, originally encoded in ICD-10-AM, was loaded to the IHPA URG grouper software version 1.4. Here the ICD-10-AM encoding was processed and a URG category

• assigned. A synthetic, constant resource value of $5,000 (a proxy National Efficient Price (NEP)) was applied

to each case, using the cost weight multiplier provided by IHPA specifications [17]. This was done to apply a constant value across all the data as different prices were applicable across these different data collection and budget years. Patient data for January to June 2016 originally encoded in SNOMED CT was first submitted to SnoMAP to achieve ICD-10-AM encoding. Any cases that remained ‘unmapped’ were curated by index hospital and re-

• submitted. After SnoMAP delivered ICD-10-AM encoded data, these cases were uploaded to the IHPA URG

grouper version 1.4, achieved a URG category and cost weights were applied as above. Additionally, AEHRC undertook a second comparative analysis where 2016 SNOMED CT AU encoded data was mapped using the previously available EDRS ( Table 2) and compared with SnoMAP processed data. Granularity of the Clinical Information

We examined the characteristics of the three data sets prior to analysis (Table 3). Data from 2014 and 2015 was originally encoded in ICD-10-AM and no mapping was undertaken. Of the 16,000 ICD-10-AM codes that were available to clinician users, less than 820 appeared in the data from these years. Data from 2016 shows clinicians entering 6007 unique SNOMED CT diagnostic terms. Here these terms were mapped (via SnoMAP) to ICD-10-AM. The ICD-10-AM encoding for this data extract shows 2,254 unique ICD-10-AM codes. These results demonstrate that more specific terminology codes captured at the point of care cannot only be meaningfully grouped into less granular categories, but also results in less homogeneous and more accurate data across these categories for reporting. The use of SNOMED CT, even when subjected to post-hoc mapping, provides richer and more descriptive data (Table 4). Table 3 Patient cases and administrative data comparison for the three data sets analysed Data characteristic 2014 2015 2016 N % N % N % Chi-square Patient cases 30234 31167 31167 NS Gender M 17155 56.74 17607 56.49 17641 56.60 F 13079 43.26 13560 43.51 13526 43.40 Triage 1 566 1.87 567 1.82 575 1.84 p<0.001(1) 2 6067 20.07 6305 20.23 6455 20.71 3 14827 49.04 15481 49.67 15848 50.85 4 6944 22.97 7007 22.48 6834 21.93 5 1840 6.09 1807 5.80 1448 4.65 Blank 7 0.02 Visit type 1 30234 100.00 31167 100.00 31127 99.87 p<0.001(2) 3 22 0.07 4 1 0.00 5 11 0.04 Blank 6 0.02 End status 1 13939 46.10 14678 47.09 17579 56.40 p<0.001(3) 2 14753 48.80 14860 47.68 12055 38.68 3 281 0.93 275 0.88 196 0.63 4 238 0.79 223 0.72 574 1.84 5 994 3.29 1100 3.53 731 2.35 6 29 0.10 31 0.10 29 0.09 Blank 3 0.01

(1)

Slight shifts from Triage category 4 & 5 (2015-15) to category 2 & 3 (2016) account for the significant result.

(2)

Visit type was not encoded in 2014-15 data, so 2016 data shows significant difference.

(3)

End status returns a significant result because there is a shift toward category 1 in 2016 data.

Table 4 Terminology and code use, and the influence on ABF outcomes (overall $ only) Data Set Characteristic 2014 2015 2016 N N N Number of unique ICD-10-AM codes 801 817 2254(a) Number of unique SNOMED CT AU concepts used na na 6007 Number of patient cases that resulted in ABF errors 1 28 Total NEP resources (Cost weight from IHPA URG grouper and specifications multiplied by $5000) per case. $25,378,031 $26,205367 $26,413,558 (a) ICD-10-AM codes achieved after SnoMAP processing Accuracy of Data for funding purposes A multivariable linear model of the data was used for estimating changes in price and hence to determine whether the year of collection was a statistically significant predictor of variability in the average price per

patient (Appendix 1). Using a linear model, the mean (price) for 2014 was $1658.71 (model intercept). The estimate for 2015 was $1646.77 (-$2.79), and for 2016 was $1642.65 (-$4.12). The other terms in the model Sex, End Status and Triage Category are included to adjust for their potential changes in relative proportions and subsequent influence on price over time. The use of SnoMAP does not distort the patient data for funding

• purposes. Funding allocations are commensurate with ABF protocols, but the patient data, encoded in

SNOMED, is now is less homogeneous, and more clinically accurate and useful. Cut down subsets such as EDRS do not perform as well as using the full SNOMED CT terminology. When 79,639 SNOMED CT terms (and not the EDRS subset) were mapped comprehensively to all valid ICD-10- AM codes (10,018), far fewer patient cases returned an ABF ‘error’ category (n=28), and thus more patient cases attract legitimate funding assigned by URG categories and the NEP. Error categories in URG assignment default to UDG. UDG cost weights are used to compensate for the lack of valid diagnostic coding, or for missing data items and provide an averaged cost weight. This accounts for the small differential price between 2016 (EDRS) data and 2016 SnoMAP data (Table 9). Patient cases that returned an error category have been

• excluded. Health economists will argue that the default to UDG cost weights should have been applied in these
• analyses. Here we are examining the direct influence of terminology and mapping of diagnoses on direct, exact,

funding outcomes; adding the UDG cost weights would have disguised the effect we are seeking to reveal. It is also true that the current ABF funding model does not include diagnoses as a fundamental driver of funding. Under these circumstances, further discussion about the veracity of URG versus UDG costings are moot. Using the same data (SNOMED encoded, 2016), and the same ABF protocols, we measured the influence of maps provided by IHPA and SnoMAP, as a way of demonstrating the impact that the scope of terminology and related maps has on data validity for funding purposes. Table 6 shows that the narrow scope of EDRS content (n=5760) and the related ICD-10-AM maps (n=3055) performs poorly under ABF protocols. Because the EDRS is essentially a short-list or a subset, the mapping is also a subset of the ICD-10-AM codes considered valid under ABF protocols. The number of patient cases that return an ABF ‘error’ category (n=5773) are the result of these short-listing and mapping constraints. Table 6 Comparison of detail (# unique ICD-10 codes) and quality of data between EDRS and SnoMAP maps.

(a) Achieved using IHPA EDRS to ICD-10-AM map (b) ICD-10-AM codes achieved after SnoMAP processing (c) Excludes ‘error’ cases that would have defaulted to a UDG cost weight.

This comparison of terminology scope and available maps shows that if the index hospital had used the IHPA AIHW endorsed maps for this data submission, they would have received 49 cents less funding per patient cases than was appropriate under ABF protocols. Table 7 compares the difference in price between using the AEHRC map and the EDRS map. There are no significant differences in price among these groups (p=0.77).

Clinical Care Impact

SnoMAP is accurate and provides an accurately coded and clinically useful diagnosis for every emergency

• visit. This rich clinical data will allow improvement of the quality and efficiency of emergency care at scale.

Conclusion

Here, we have described the first validated tool for automated mapping of SNOMED CT AU to ICD10 AM. We also describe clinicians adopting a rich vocabulary of thousands of descriptive and specific terms to document patient care. This change in clinician behaviour provides a counter-point to the traditional pervasive view that clinician users will only tolerate a cut down, short-list of clinical terms. Mapping, as performed through SnoMAP, is a viable approach to ensuring that hospitals using SNOMED CT can continue to meet Data Set Characteristics 2016 EDRS treatment 2016 SnoMAP treatment Number of unique ICD-10-AM codes 1277(a) 2254(b) Number of unique SNOMED CT AU concepts used 6007 6007 Number of patient cases that resulted in ABF errors 5773 28 Total NEP resources (Cost weight from IHPA URG grouper and specifications, multiplied by $5000 per case). $20,391,038(c) $26,413,558(c)

their statutory reporting. . However, maps are not a long-term solution. Maps have to be constantly maintained and enhanced, with commensurate ongoing costs. SnoMAP can be widely used to bridge the divide between clinically useful code sets in hospitals and administrative requirements. This flexibility will allow retention of funding, while freeing clinicians to code accurately to develop databases to accurately improve the quality and efficiency of care they provide.

References

• 1. Sullivan C, Staib A, Ayre S, et al. Pioneering digital disruption: Australia's first integrated digital tertiary
• hospital. The Medical journal of Australia 2016;205(9):386-89
• 2. Australian Institute of Health and Welfare. Australian Institute of Health and Welfare: Australian hospital

statistics 2013–14: emergency department care. Health services series no. 58. Cat. no. HSE 153. Canberra: AIHW. Secondary Australian Institute of Health and Welfare: Australian hospital statistics 2013–14: emergency department care. Health services series no. 58. Cat. no. HSE 153. Canberra: AIHW. 2014 2014. http://www.aihw.gov.au/publication-detail/?id=60129549039.

• 3. Australian Institute of Health and Welfare. Non-admitted patient emergency department care NMDS 2013-
• 14. Secondary Non-admitted patient emergency department care NMDS 2013-14 2014.

http://meteor.aihw.gov.au/content/index.phtml/itemId/509116.

• 4. Independent Hospital Pricing Authority. Emergency Care. Secondary Emergency Care 2016.

https://www.ihpa.gov.au/what-we-do/emergency-care.

• 5. Independent Hospital Pricing Authority. Activity Based Funding. Secondary Activity Based Funding 2016.

https://www.ihpa.gov.au/what-we-do/activity-based-funding.

• 6. Australian Consortium for Classification Development. ICD-10-AM/ACHI/ACS. Secondary ICD-10-

AM/ACHI/ACS 2014. https://www.accd.net.au/Icd10.aspx. 7. Agency ADH. Clinical Terminology. Secondary Clinical Terminology 2017. http://www.digitalhealth.gov.au/get-started-with-digital-health/what-is-digital-health/clinical- terminology.

• 8. Stausberg J, Lehmann N, Kaczmarek D, Stein M. Reliability of diagnoses coding with ICD-10. Int J Med

Inform 2008;77(1):50-7 doi: 10.1016/j.ijmedinf.2006.11.005[published Online First: Epub Date]|.

• 9. National Center for Health Statistics. National Vital Statistics System. Secondary National Vital Statistics

System 2015. https://www.cdc.gov/nchs/nvss/deaths.htm.

• 10. France R. Case Mix: Global Views, Local Actions: Evolution in Twenty Countries. Amsterdam: IOS Press,

2001.

• 11. World Health Organisation. Classification of Diseases. Secondary Classification of Diseases 2016.

http://www.who.int/classifications/icd/en/.

• 12. Australian Institute of Health and Welfare. Admitted patient care NMDS 2013-14. Secondary Admitted

patient care NMDS 2013-14 2014. http://meteor.aihw.gov.au/content/index.phtml/itemId/491555.

• 13. Sullivan C, Staib A, Khanna S, et al. The National Emergency Access Target (NEAT) and the 4-hour rule:

time to review the target. The Medical journal of Australia 2016;204(9):354

• 14. Independent Hospital Pricing Authority. Technical Specifications 2015-16 National Pricing Model.

Secondary Technical Specifications 2015-16 National Pricing Model 2015. https://www.ihpa.gov.au/sites/g/files/net636/f/publications/national_pricing_model_technical_spe cifications_2015-16.pdf.

• 15. Hansen DP, Kemp ML, Mills SR, Mercer MA, Frosdick PA, Lawley MJ. Developing a national emergency

department data reference set based on SNOMED CT. The Medical journal of Australia 2011;194(4):S8-10

• 16. National Health Service. Terminology and Classifications: National standards for recording and

categorising information to support care delivery, statistical analysis, research and the reimbursement of health and care providers. Secondary Terminology and Classifications: National standards for recording and categorising information to support care delivery, statistical analysis, research and the reimbursement

• f

health and care providers 2017. https://digital.nhs.uk/article/290/Terminology-and-Classifications.

• 17. independent Hospital Pricing Authority. Investigative review of classification systems for emergency care.

Secondary Investigative review of classification systems for emergency care 2014. https://www.ihpa.gov.au/sites/g/files/net636/f/publications/classification-systems-emergency- care-v4.0.pdf.