Table 6.
Performance of various models and traditional ICD code approaches to define and identify patients with HF among 200 patients with and without gold standard HF where gold standard HF that excluded possible HF
| Gold standard HF | HF classified by models or ICD codes | Number of gold standard HF within the HF cohort | Precision (PPV) | Recall (Sensitivity) | F score | Accuracy | |
|---|---|---|---|---|---|---|---|
| 1. NLP | 80 | 52 | 40 | 76.9% | 50.0% | 69.5% | 74.0% |
| 2. ML | 80 | 73 | 53 | 72.6% | 66.3% | 71.5% | 77.0% |
| 3. NLP + ML | 80 | 64 | 51 | 79.6% | 63.7% | 76.1% | 79.0% |
| 4. ICD codes | 80 | ||||||
| a: One or more principal hospital discharge diagnosis | 80 | 31 | 26 | 83.9% | 32.5% | 63.7% | 70.5% |
| b. Two or more primary outpatient encounter diagnoses | 80 | 89 | 53 | 59.6% | 66.3% | 60.8% | 68.5% |
| c. One principal hospital discharge diagnosis or two primary outpatient encounter diagnoses | 80 | 91 | 55 | 60.4% | 68.8% | 61.9% | 69.5% |
HF, heart failure; ICD, International Classification of Diseases; ML, machine learning; NLP, natural language processing; PPV, positive predictive value.