Diagnostic accuracy of existing methods for identifying diabetic foot ulcers from inpatient and outpatient datasets
© Sohn et al; licensee BioMed Central Ltd. 2010
Received: 6 October 2010
Accepted: 24 November 2010
Published: 24 November 2010
As the number of persons with diabetes is projected to double in the next 25 years in the US, an accurate method of identifying diabetic foot ulcers in population-based data sources are ever more important for disease surveillance and public health purposes. The objectives of this study are to evaluate the accuracy of existing methods and to propose a new method.
Four existing methods were used to identify all patients diagnosed with a foot ulcer in a Department of Veterans Affairs (VA) hospital from the inpatient and outpatient datasets for 2003. Their electronic medical records were reviewed to verify whether the medical records positively indicate presence of a diabetic foot ulcer in diagnoses, medical assessments, or consults. For each method, five measures of accuracy and agreement were evaluated using data from medical records as the gold standard.
Our medical record reviews show that all methods had sensitivity > 92% but their specificity varied substantially between 74% and 91%. A method used in Harrington et al. (2004) was the most accurate with 94% sensitivity and 91% specificity and produced an annual prevalence of 3.3% among VA users with diabetes nationwide. A new and simpler method consisting of two codes (707.1× and 707.9) shows an equally good accuracy with 93% sensitivity and 91% specificity and 3.1% prevalence.
Our results indicate that the Harrington and New methods are highly comparable and accurate. We recommend the Harrington method for its accuracy and the New method for its simplicity and comparable accuracy.
With the rapid spread of electronic medical records, there is a growing need for accurately identifying health conditions through electronic medical records in order to establish population-based rates for disease surveillance purposes and to cost-effectively identify patients for targeted interventions and research studies. Diabetic foot ulcers (DFUs) are significant public health concerns due to high economic burden [1–4], negative impact on quality of life [5, 6], and their association with increased risk of amputation [7, 8] and premature death [9, 10]. However, their national estimates of incidence or prevalence rates are not currently available, possibly due to the lack of a reliable method to identify this condition in administrative health data. We only know that a lifetime risk of foot ulceration for a diabetic patient may be as high as 25%  and that annual incidence and prevalence rates may be as high as 4% and 10% in selected populations [12, 13].
Four different methods [1–3, 14] have been used in previous observational studies. They differed considerably from one another in complexity and sophistication; they were designed for different purposes and were used with different databases. In a study of costs and duration of treatment for foot ulcer patients, Holzer and colleagues  identified DFU patients from inpatient and outpatient claims data. Any patient with one or more claims containing a foot ulcer-related diagnosis or procedure in any fields was identified as having the DFU diagnosis.
In a descriptive study of inpatient care for patients with lower-extremity complications of diabetes, Mayfield et al.  reported that over 18,000 hospitalizations for lower-extremity complications occurred in 1998. They identified foot ulcers using a method consisting of diagnostic codes only. Venous stasis ulcers and decubitus ulcers were excluded but surgical complications from a stump infection, an orthopaedic procedure, or a prior vascular graft in the foot were identified as a DFU.
Ramsey et al. [3, 15] used the simplest method, involving only one diagnostic code (ICD-9-CM 707.1×, "Ulcer of lower limbs, except decubitus"), in a study of incidence rates and treatment costs of foot ulcers among individuals enrolled in a HMO. In a validation study, this method was shown to have 74% sensitivity and 94% specificity compared to medical records .
Existing methods of identifying diabetic foot ulcers in administrative data
ICD-9-CM or CPT-4 codes
A. Lower-extremity ulcer diagnosis
Ulcer of lower limbs
Chronic ulcer of other specified sites
Chronic ulcer of unspecified sites
Carbuncle and furnancle of foot
Cellulitis and abscess of toe or foot
Cellulitis and abscess of unspecified digit
Other cellulitis and abscess, leg except foot
730.06-730.09, 730.16-730.19, 730.26-730.29
Surgical complications from a stump infection
Surgical complications from amputation
Complications from a prior vascular graft
440.3, 996.62, 996.7, 996.74, E878.2
B. Lower-extremity ulcer-related procedures
Simple repair of superficial wound
11040-11044, 77.68, 86.22, 86.28
Surgical debridement and drainage of abscess and cavities
Lower-extremity radiographic techniques
Angioscopy, arteriography, angiography
Lower-extremity CAT or MRI scanning
Incision or excision of foot
Unna boot application
29540, 29550, 29580
Culture and sensitivity testing
87040, 87071-87072, 87075-87076, 87082-87085
Aspiration, incision and drainage of infection or abscess
10060-10061, 10160, 20000, 86.01, 86.04
28020-28024, 28060, 28070, 28072, 28086, 28088, 28110-28126, 28140, 28150, 28153, 28160, 77.38, 77.88, 80.18
Late amputation stump complication5
Amputation, knee and above
The objectives of this study were to compare these four methods for their diagnostic accuracy by evaluating them using medical records as the gold standard and to propose a new and simpler method.
Study cohort and data sources
To evaluate the diagnostic coding accuracy of these methods, we first identified all individuals who used the Department of Veterans Affairs (VA) healthcare services in the fiscal year 2003 (October 1, 2002-September 31, 2003; all years hereafter are fiscal years) from the VA national patient care datasets. These datasets contain all records of acute inpatient or outpatient care provided in the US. Patients were identified as having diabetes if they received at least one prescription for a diabetes medication in the current year or if two or more records with diabetes diagnosis (ICD-9-CM 250.xx) existed for inpatient admissions or outpatient visits over a 24-month period (2002-2003). This method is known to have 93% sensitivity and 98% specificity relative to self reports of diabetes .
From the national diabetic cohort (N = 866,881), we identified all patients who used healthcare services exclusively at a tertiary care hospital in 2003. We identified 4,158 diabetic patients from whom we drew a stratified sample consisting of all individuals who had DFUs according to at least one of the four methods and an equal number of individuals who were randomly selected from those who did not. This resulted in a hospital-based sample of 518 individuals, which we will call the "local" sample below.
Review of medical records
We provided two authors (EB and FS) with a list of 518 individuals that did not have any indication of whether a diagnosis of a foot ulcer was found in administrative data. EB and FS divided the list into half and independently reviewed patients' electronic medical records. Their aim was to determine whether a diabetic foot ulcer was indicated on medical records in 2003. A diabetic foot ulcer was conceptually defined as a full-thickness break of the integument on a diabetic foot. It was indicated if there was any explicit mention of "diabetic foot ulcer" or any qualifying wound or lesion on an ankle or a foot was noted on medical records. When osteomyelitis or gangrene was mentioned alone in 2003, we identified it as a DFU if we could link it to foot ulceration on the same foot and location in 2002. Osteomyelitis due to puncture wounds, gangrene due to arterial occlusion/embolic phenomenon, abrasions, venous stasis ulcers, and decubitus ulcers were excluded from the case definition.
There were 45 cases whose DFU status could not be unambiguously determined by the reviewers. These cases were examined by both EB and FS and a third reviewer (RS). When there were disagreements between EB and FS, we used the opinion of the third reviewer to adjudicate the case. To assess inter-rater reliability, we randomly selected 30 medical records de novo from the "local" sample and all three reviewers (EB, FS, RS) independently conducted the reviews. Cronbach's alpha for the inter-rater reliability among three reviewers was 0.93, indicating a high consistency.
New identification method
In addition to evaluating existing methods, we developed a new, simple method for DFU identification. The New method consisted of two codes 707.1× and 707.9 documented in any position on an inpatient or outpatient encounter. These two codes were common to the Holzer, Mayfield, and Harrington methods and thus the New method will identify a subset of patients also identified by the first three methods.
Foot ulcer indication in medical charts was used as the "gold standard" against which four methods were evaluated for diagnostic accuracy. Sensitivity and specificity were computed for each method. Sensitivity indicates the probability that a foot ulcer indication on medical charts is correctly identified by a method. Specificity indicates the probability that a patient who does not have an indication on medical charts is not identified as having the condition by a method. We additionally computed weighted positive predictive value (PPV) and negative predictive value (NPV) to account for disproportionate sampling in the "local" sample . PPV indicates the proportion of patients a method correctly predicts a foot ulcer indication on medical records and NPV, the proportion a method correctly excludes as not having a foot ulcer indication on medical records. Simple kappa, weighted to adjust for bias due to disproportionate sampling, was computed for each method as a measure of agreement between administrative data and medical charts [18, 19]. Sampling weights used for PPV, NPV, and kappa were the inverse of the probability of selection to the local sample.
The study was approved by the Institutional Review Board at the Hines VA Hospital.
Prevalence rates of diabetic foot ulcers based on four methods
We identified 866,881 patients who used VA healthcare services in the US in 2003 with a diagnosis of diabetes. They were 68 ± 11 years old, mostly male (98%) and non-Hispanic whites (71%). Sixteen percent were newly diagnosed with diabetes in 2003 and 24% had had diabetes for 6 years or longer.
Diabetic foot ulcer prevalence according to five methods (N = 866,881)
Agreement among methods*
A comparison among methods shown in Table 2 suggests that Holzer and Mayfield methods identified essentially all patients who were also identified by the other two methods. All other methods captured 100% of those who were identified by the Ramsey method, indicating that the Ramsey method was the least common denominator of all methods.
Comparison of accuracy
Comparison of methods for diagnostic accuracy of diabetic foot ulcers (N = 518)
Accuracy and agreement measures (95% CI)†
The Ramsey method was similar in all measures to the Harrington method, but the former can capture only 83% of DFU patients identified by the latter in the national diabetic population as shown in Table 1. In contrast, the Ramsey method produced the smallest number incorrectly classified (43 false positives plus true negatives, 8.3% of the local sample), followed by the Harrington method with 45 (8.7%). The other two methods fared worse with 67 for the Holzer (12.9%) and 99 (19.1%) for the Mayfield method.
We found that a fifth method ("New" in Tables 2 and 3) that consisted of two codes 707.1× and 707.9 performed as well as the Harrington method with 92.9% sensitivity and 90.9% specificity and 44 (8.5%) incorrectly classified. Kappa for the New method was 0.73, indicating substantial agreement with medical records .
Our objective in this study was to evaluate diagnostic coding accuracy of four existing methods compared to medical records. We showed that the five methods we examined in this study performed very well in sensitivity. Holzer and Mayfield methods identified a large number of false positives with a resulting low specificity and positive predictive values. The last three methods (Ramsey, Harrington, and New) had sensitivity > 92% for coding accuracy and were similar in specificity (90.1-91.4), even though the number of diagnostic and procedure codes involved varied considerably. We also showed that the DFU prevalence based on five methods varied considerably. The Mayfield method identified 41% more cases than the Ramsey method, suggesting that the choice of a method can substantially influence prevalence estimates.
As far as we know, the Ramsey method was the only one that was previously evaluated for accuracy. Compared with medical records for patients enrolled in a commercial healthcare plan, this method had 74% sensitivity and 94% specificity . A study by Harwell et al.  evaluated an algorithm for "foot complications" that included DFUs, Charcot arthropathy, and lower-extremity revascularization or bypass procedures. Their algorithm was based on the Harrington method (for identifying DFUs that comprise the large majority of foot complications) with additional codes for Charcot arthropathy and lower-extremity vascular procedures. This algorithm had excellent accuracy (99% sensitivity and 93% specificity) in identifying foot complications from inpatient administrative records. These results are consistent with ours on the Harrington method, even though sensitivity and specificity are much higher in the Harwell et al. study than in ours. The difference may be attributed to the fact that the results from the Harwell et al. study were obtained from inpatient administrative records and ours from both inpatient and outpatient records, and to the fact that their case definition is much broader ("foot complications") than ours (DFUs).
This study has limitations. The measures of agreement for different methods in this study may not be generalizable to non-VA databases to the extent that the practices for coding foot ulcers are different from system to system. In principle, the VA uses coding guidelines that are also used in the rest of the medical community, namely, the Official Guidelines for Coding and Reporting approved by the American Hospital Association, the American Health Information Management Association, the Centers for Medicare and Medicaid Services, and the National Center for Health Statistics . Variation in adherence to these guidelines, coding intensity, and data quality among providers need to be considered when applying the results of this study to non-VA data such as Medicare claims. Further research is also needed to confirm whether our findings based on the VA data can be applied to the non-VA data.
Another limitation is that the disease coding in the administrative data were not matched with medical charts kept on the same date. It was not practicable for us to match every eligible code used in Harrington or Holzer methods with medical charts for the same date. Establishing the accuracy of diagnostic coding for each administrative health record is important for determining, for example, the first date of diagnosis or whether a disease existed before or after the onset of another disease. In a supplemental analysis, we assessed the accuracy at the code-day level by randomly selecting 30 patients with encounters coded with 707.1× or 707.9 in the local sample and matched their encounters with medical charts for the same date. We found that 29 (97%) were corroborated by medical charts, suggesting an excellent accuracy of the New method at the code-day level in the VA data.
Our chart reviews show that administrative data can be used to identify persons with DFU with considerably higher accuracy than previously believed. The accuracy of DFU identification can be as high as some of the high-risk, high-profile conditions that have received a lot of research and policy attention such as myocardial infarction. Our results indicate that the Harrington and New methods are highly comparable and accurate. We recommend the Harrington method for its accuracy and the New method for its simplicity and comparable accuracy. The Harrington method showed 94% sensitivity and 90% specificity in accuracy in the VA administrative data. According to this method, the annual prevalence of diabetic foot ulcers was 3.3% in the VA diabetic population in 2003.
List of abbreviations
Diabetic foot ulcers
negative predictive value
positive predictive value
The Department of Veterans Affairs
The authors gratefully acknowledge the financial support from the Center for Management of Complex Chronic Care, Hines VA Hospital, Hines, IL (LIP 42-522; Elly Budiman-Mak, MD, Principal Investigator). The paper presents the findings and conclusions of the authors; it does not necessarily represent the Department of Veterans Affairs or Health Services Research and Development Service. We are also grateful to Dr. Julia Riley for her initial work on chart reviews. The corresponding author had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
- Harrington C, Zagari MJ, Corea J, Klitenic J: A cost analysis of diabetic lower-extremity ulcers. Diabetes Care. 2000, 23: 1333-1338. 10.2337/diacare.23.9.1333.PubMedGoogle Scholar
- Holzer SE, Camerota A, Martens L, Cuerdon T, Crystal-Peters J, Zagari M: Costs and duration of care for lower extremity ulcers in patients with diabetes. Clin Ther. 1998, 20: 169-181. 10.1016/S0149-2918(98)80044-1.PubMedGoogle Scholar
- Ramsey SD, Newton K, Blough D, McCulloch DK, Sandhu N, Reiber GE, Wagner EH: Incidence, outcomes, and cost of foot ulcers in patients with diabetes. Diabetes Care. 1999, 22: 382-387. 10.2337/diacare.22.3.382.PubMedGoogle Scholar
- Ramsey SD, Newton K, Blough D, McCulloch DK, Sandhu N, Wagner EH: Patient-level estimates of the cost of complications in diabetes in a managed-care population. Pharmacoeconomics. 1999, 16: 285-295. 10.2165/00019053-199916030-00005.PubMedGoogle Scholar
- Nabuurs-Franssen MH, Huijberts MS, Nieuwenhuijzen Kruseman AC, Willems J, Schaper NC: Health-related quality of life of diabetic foot ulcer patients and their caregivers. Diabetologia. 2005, 48: 1906-1910. 10.1007/s00125-005-1856-6.PubMedGoogle Scholar
- Armstrong DG, Lavery LA, Wrobel JS, Vileikyte L: Quality of life in healing diabetic wounds: does the end justify the means?. J Foot Ankle Surg. 2008, 47: 278-282. 10.1053/j.jfas.2008.02.015.PubMedGoogle Scholar
- Adler AI, Boyko EJ, Ahroni JH, Smith DG: Lower-extremity amputation in diabetes. The independent effects of peripheral vascular disease, sensory neuropathy, and foot ulcers. Diabetes Care. 1999, 22: 1029-1035. 10.2337/diacare.22.7.1029.PubMedGoogle Scholar
- Mayfield JA, Reiber GE, Maynard C, Czerniecki JM, Caps MT, Sangeorzan BJ: Trends in lower limb amputation in the Veterans Health Administration, 1989-1998. J Rehabil Res Dev. 2000, 37: 23-30.PubMedGoogle Scholar
- Boyko EJ, Ahroni JH, Smith DG, Davignon D: Increased mortality associated with diabetic foot ulcer. Diabet Med. 1996, 13: 967-972. 10.1002/(SICI)1096-9136(199611)13:11<967::AID-DIA266>3.0.CO;2-K.PubMedGoogle Scholar
- Moulik PK, Mtonga R, Gill GV: Amputation and mortality in new-onset diabetic foot ulcers stratified by etiology. Diabetes Care. 2003, 26: 491-494. 10.2337/diacare.26.2.491.PubMedGoogle Scholar
- Singh N, Armstrong DG, Lipsky BA: Preventing foot ulcers in patients with diabetes. JAMA. 2005, 293: 217-228. 10.1001/jama.293.2.217.PubMedGoogle Scholar
- Lavery LA, Armstrong DG, Wunderlich RP, Tredwell J, Boulton AJ: Diabetic foot syndrome: evaluating the prevalence and incidence of foot pathology in Mexican Americans and non-Hispanic whites from a diabetes disease management cohort. Diabetes Care. 2003, 26: 1435-1438. 10.2337/diacare.26.5.1435.PubMedGoogle Scholar
- Singh N, Armstrong DG, Lipsky BA: Preventing foot ulcers in patients with diabetes. JAMA. 2005, 293: 217-228. 10.1001/jama.293.2.217.PubMedGoogle Scholar
- Mayfield JA, Reiber GE, Maynard C, Czerniecki J, Sangeorzan B: The epidemiology of lower-extremity disease in veterans with diabetes. Diabetes Care. 2004, 27 (Suppl 2): B39-B44. 10.2337/diacare.27.suppl_2.B39.PubMedGoogle Scholar
- Newton KM, Wagner EH, Ramsey SD, McCulloch D, Evans R, Sandhu N, Davis C: The use of automated data to identify complications and comorbidities of diabetes: a validation study. J Clin Epidemiol. 1999, 52: 199-207. 10.1016/S0895-4356(98)00161-9.PubMedGoogle Scholar
- Miller DR, Safford MM, Pogach LM: Who has diabetes? Best estimates of diabetes prevalence in the Department of Veterans Affairs based on computerized patient data. Diabetes Care. 2004, 27 (Suppl 2): B10-B21. 10.2337/diacare.27.suppl_2.B10.PubMedGoogle Scholar
- Altman DG, Bland JM: Diagnostic tests 2: Predictive values. BMJ. 1994, 309: 102-PubMedPubMed CentralGoogle Scholar
- Craig BM, Adams AK: Accuracy of body mass index categories based on self-reported height and weight among women in the United States. Matern Child Health J. 2009, 13: 489-496. 10.1007/s10995-008-0384-7.PubMedGoogle Scholar
- Chen G, Faris P, Hemmelgarn B, Walker RL, Quan H: Measuring agreement of administrative data with chart data using prevalence unadjusted and adjusted kappa. BMC Med Res Methodol. 2009, 9: 5-10.1186/1471-2288-9-5.PubMedPubMed CentralGoogle Scholar
- Landis JR, Koch GG: The measurement of observer agreement for categorical data. Biometrics. 1977, 33: 159-174. 10.2307/2529310.PubMedGoogle Scholar
- Harwell TS, Gilman J, Dehart L, Loran E, Eyler N, Schrumpf P, Corsi CM, McDowall JM, Johnson EA, Ford JA, et al: Validation of a case definition for foot complications among hospitalized patients with diabetes. Diabetes Care. 2002, 25: 630-631. 10.2337/diacare.25.3.630.PubMedGoogle Scholar
- ICD-9-CM Official Guidelines for Coding and Reporting. 2006, Accessed on December 3, 2010, [http://www.cdc.gov/nchs/data/icd9/icdguide09.pdf]
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (<url>http://creativecommons.org/licenses/by/2.0</url>), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.