Updating The Accuracy of Administrative Claims for Identifying Left Ventricular Ejection Fraction Among Patients with Heart Failure

Left ventricular ejection fraction (EF) is important in determining guideline-directed heart failure (HF) therapy. However, EF is unavailable in administrative claims, limiting this data for quality measures or research. Previous analyses using International Classification of Disease-Ninth Revision (ICD-9) codes demonstrated low sensitivity for identifying HF patients with reduced (rEF) or preserved ejection fraction (pEF).^1-3 We used the Veteran’s Affairs (VA) EF natural language processing algorithm to evaluate the predictive accuracy of ICD-10 HF codes.

METHODS

Study data are available to VA researchers; the analysis code will be provided on request. We identified HF diagnoses for VA patients between 2017-2019 from VA, non-VA fee care, and Medicare claims. We leveraged a natural language processing algorithm with >95% precision to extract EF from clinical notes and imaging reports.⁴ We excluded EF with ranges exceeding 10% as potential errors; this removed 5.3% of EF estimates.

For each diagnosis, we identified the closest EF within 180 days. We determined the proportion with EF≤40%, 40-50%, or ≥50% across codes (Table 1). We classified codes as HFrEF-related if over half had EF≤40% and HFpEF-related if over half had EF≥50%. Codes meeting neither criterion or with frequency <1,000 were termed non-specific.

Table 1.

Performance of HF Diagnosis Claims for Classifying Ejection Fraction

Diagnosis-Level Analysis
Diagnosis*	Code	Count	EF≤40%	EF 40-50%	EF≥ 50%	Classification
All Codes		1,671,084	45.8%	13.5%	40.7%	-
Left Ventricular Failure Unspecified	I50.1	12,007	43.8%	16.8%	39.5%	Non-specific
Systolic HF	I50.2X	414,989	69.3%	15.5%	15.2%	rEF
Diastolic HF	I50.3X	281,254	13.4%	8.9%	77.7%	pEF
Combined Systolic and Diastolic HF	I50.4X	100,733	59.4%	18.0%	22.5%	rEF
Right HF	I50.810-I50.813	6,662	18.9%	9.3%	71.8%	pEF
Biventricular HF	I50.82	4,534	70.1%	11.4%	18.5%	rEF
End-stage HF	I50.84	3,462	85.7%	4.6%	9.6%	rEF
Other HF	I50.89-I50.9	295,380	43.1%	13.2%	43.6%	Non-specific
Hypertensive Heart Disease with HF	I11.0, I13.0, I13.2	551,503	43.3%	13.8%	42.9%	Non-specific

Patient-Level Analysis (n=274,202)
Predictor	% Classified	EF	AUC	Threshold >0		Threshold ≥0.5		Threshold 1
				Sn	PPV	Sn	PPV	Sn	PPV
(1) Proportion of Specific HF Codes†	76.7%	≤40%	0.76	91.2%	59.5%	90.3%	61.1%	83.4%	63.2%
		≤45%	0.79	89.9%	71.6%	88.7%	73.3%	81.4%	75.3%
		<50%	0.81	88.6%	77.1%	87.3%	81.2%	79.9%	83.0%
(2) Proportion of Total Codes†	100.0%	≤40%	0.73	75.1%	59.5%	59.2%	62.7%	15.0%	59.6%
		≤45%	0.75	73.4%	71.6%	57.5%	74.8%	14.4%	70.8%
		<50%	0.75	72.0%	77.1%	56.1%	80.2%	14.2%	76.3%
Sensitivity Analyses‡
Scenario	% Classified	EF	AUC	Threshold >0		Threshold ≥0.5		Threshold 1
				Sn	PPV	Sn	PPV	Sn	PPV
E&M & Principal Inpatient Diagnoses (n=115,001)	53.5%	≤40%	0.75	91.9%	66.2%	91.3%	67.0%	87.6%	68.3%
		<50%	0.80	89.4%	82.0%	88.6%	82.7%	84.5%	83.9%
Principal Inpatient Diagnosis Alone (n=71,932)	39.6%	≤40%	0.77	91.4%	71.8%	---	---	---	---
		<50%	0.82	88.6%	86.2%	---	---	---	---
EF Within 30 Days of Diagnosis (n=223,816)	80.5%	<40%	0.77	92.1%	61.2%	91.0%	63.1%	83.4%	65.6%
		<50%	0.82	89.8%	78.9%	88.3%	80.9%	80.1%	83.2%
≥4 Diagnoses in Prior Year (n=75,253)	94.6%	≤40%	0.79	94.9%	61.8%	92.5%	66.3%	77.7%	72.1%
		<50%	0.85	93.2%	78.3%	89.7%	82.9%	73.3%	87.8%
Diagnoses from Prior 90 Days (n=274,202)	73.6%	≤40%	0.76	90.3%	60.7%	90.0%	61.4%	86.2%	62.7%
		<50%	0.80	87.5%	78.4%	87.0%	79.1%	82.9%	80.3%

^*Codes not listed (I50.814, Right Heart Failure due to Left HF; I50.83, High Output HF; I09.81, Rheumatic HF) were classified as non-specific due to total count <1,000 in the database.

^†Classification based on the proportion of specific HF codes is defined as the number of HFrEF-specific codes divided by the number of HF codes excluding non-specific codes for each patient. Classification based on the proportion of total codes is based on the number of HFrEF-specific codes divided by the total number of HF codes for each patient. Prediction performance is assessed at thresholds of 0, 0.5, and 1.

^‡Sensitivities for predictor (1): proportion of specific HF codes classified as HFrEF-related; Abbreviations: AUC indicates area under the curve; EF, ejection fraction; E&M, evaluation and management; pEF, preserved ejection fraction; PPV, positive predictive value; rEF, reduced ejection fraction; Sn, sensitivity.

To test EF classification using multiple diagnoses, we identified a random diagnosis between 2018-2019 for each patient and all HF diagnoses in the prior year. We evaluated two patient-level predictors: (1) the proportion of specific HF diagnoses classified as HFrEF-related and (2) the proportion of all HF diagnoses classified as HFrEF-related. We assessed three thresholds for identifying HFrEF: >0 (i.e., any HFrEF diagnosis), ≥0.5, and 1 (i.e., all HFrEF diagnoses). We calculated the area under the curve (AUC), sensitivity, and positive predictive value (PPV) for identifying EF≤40%, ≤45%, and <50%.

We performed multiple sensitivity analyses: (1) only evaluation and management or inpatient principal diagnoses, (2) inpatient principal diagnosis alone, (3) EF within 30 days, (4) patients with ≥4 diagnoses, and (5) diagnoses within prior 90 days. This was approved by the Stanford Institutional Review Board; data is available to VA researchers.

RESULTS

Between 2017-2019, we identified 11,817,035 HF diagnoses across 993,408 individuals. There were 358,172 patients and 1,671,084 diagnoses with an EF within 180 days. This included 398,650 (23.9%) VA outpatient, 652,716 (39.1%) VA inpatient, 279,729 (16.7%) non-VA outpatient, and 339,989 (20.3%) non-VA inpatient diagnoses. Median duration between diagnosis and EF was 1 day (IQR: 1-14 days). Median EF was 43% (IQR: 30-55%).

Table 1 lists the EF subgroup breakdown (EF≤40%, 40-50%, ≥50%) for each diagnosis. Among the 523,718 diagnoses classified as HFrEF, 67.6% had EF≤40% compared with 16.6% with EF≥50%. Among the 287,916 diagnoses classified as HFpEF, 77.6% had EF≥50% compared with 13.5% with EF≤40%. There were 859,450 non-specific diagnoses.

Median time between EF and diagnosis was shorter among VA diagnoses than non-VA diagnoses (1 day [IQR 0-1] vs. 17 days [IQR 5-52]). VA diagnoses classified as HFrEF were more likely to have EF≤40% compared with non-VA diagnoses (68.9% vs. 64.8%, p<0.01).

We identified a random diagnosis for 274,202 patients between 2018-2019. The average age was 74.0 (SD: 10.5) with 2.6% women. This included 63.5% non-Hispanic White , 5.0% Hispanic, and 15.8% non-Hispanic Black with 13.8% with unknown race/ethnicity. 75.4% had coronary artery disease, 43.9% chronic kidney disease, and 58.2% diabetes.

Table 1 displays predictor performance. Predictor 1 - proportion of specific diagnoses classified as HFrEF-related – had an AUC for predicting EF≤40% of 0.76, which increased to 0.80 for EF<50%. 23.3% of patients with only non-specific HF diagnoses were not characterized. Using the proportion of all diagnoses classified as HFrEF-related (predictor 2) enabled predictions across the cohort with decreased AUC of 0.73. Predictor 1 performed better among patients with ≥4 diagnoses in the prior year (AUC 0.79 for EF≤40%). At a threshold of ≥1 HFrEF-related diagnosis, sensitivity was 94.9% and PPV was 61.8%. Requiring all diagnoses to be HFrEF-related increased specificity to 72.1% but decreased sensitivity to 77.7%.

DISCUSSION

Among Veterans, administrative claims had moderate accuracy (AUC 0.76) at identifying HFrEF using the proportion of specific HF diagnoses classified as HFrEF-related. However, a quarter of patients had only non-specific diagnoses.

HFrEF classification improved with an EF threshold of <50% because clinicians frequently use systolic dysfunction codes for mid-range EF. However, quality measures typically focus on EF≤40. Among patients classified as HFrEF, 15-20% have an EF 40-50% and 20-25% have EF>50%. Incorporating other characteristics may improve classification, but including prior treatment or comorbidities may bias quality measures.

There are limitations in generalizing these results to alternate populations. The PPV may decrease among populations with a higher HFpEF prevalence.

Current HF diagnosis codes remain inadequate for defining populations for quality measures or for comparative effectiveness research. However, claims-based EF algorithms may be acceptable for capturing patients with EF<50%. Administrative codes require better alignment with clinical definitions. Fortunately, ICD-11 has EF-specific diagnoses. Until then, studying HF with claims data will remain challenging.