Table 2.
Main results on DRG prediction.
| DRG set | MACRO-AUC | MICRO-AUC | MACRO-F1 | MICRO-F1 | Number of DRG targets | Number (%) of hospital stays | |
|---|---|---|---|---|---|---|---|
| MS-DRG | |||||||
| CAML with clinical text | All DRGs | 0·871 (0·011) | 0·956 (0·002) | 0·084 (0·008) | 0·270 (0·006) | 369 | 1977 (100·0) |
| MDC 05 | 0·836 (0·016) | 0·974 (0·002) | 0·141 (0·010) | 0·382 (0·011) | 67 | 522 (26·4) | |
| MDC 01 | 0·850 (0·011) | 0·974 (0·001) | 0·121 (0·008) | 0·343 (0·016) | 52 | 320 (16·2) | |
| MDC 04 | 0·833 (0·013) | 0·971 (0·003) | 0·112 (0·018) | 0·294 (0·024) | 34 | 219 (11·1) | |
| Top 80% cases | 0·923 (0·005) | 0·983 (0·001) | 0·192 (0·012) | 0·329 (0·008) | 131 | 0·923 (0·005) | |
| Top 50 DRGs | 0·940 (0·002) | 0·991 (0·000) | 0·320 (0·014) | 0·436 (0·009) | 50 | 1083 (54·8) | |
| Top 30 DRGs | 0·943 (0·002) | 0·993 (0·001) | 0·395 (0·013) | 0·502 (0·006) | 30 | 842 (42·6) | |
| LSTM with clinical measurements | All DRGs | 0·819 (0·009) | 0·940 (0·004) | 0·041 (0·008) | 0·183 (0·011) | 369 | 1977 (100·0) |
| APR-DRG | |||||||
| CAML with clinical text | All DRGs | 0·884 (0·003) | 0·963 (0·001) | 0·069 (0·008) | 0·244 (0·005) | 517 | 2747 (100·0) |
| MDC 05 | 0·881 (0·007) | 0·986 (0·000) | 0·116 (0·004) | 0·333 (0·006) | 87 | 733 (26·7) | |
| MDC 01 | 0·892 (0·004) | 0·984 (0·001) | 0·123 (0·019) | 0·321 (0·019) | 63 | 424 (15·4) | |
| MDC 04 | 0·759 (0·012) | 0·963 (0·002) | 0·096 (0·032) | 0·193 (0·016) | 47 | 319 (11·6) | |
| Top 80% cases | 0·936 (0·002) | 0·985 (0·000) | 0·180 (0·013) | 0·306 (0·004) | 180 | 2197 (80·0) | |
| Top 50 DRGs | 0·952 (0·001) | 0·993 (0·000) | 0·336 (0·005) | 0·434 (0·008) | 50 | 1285 (46·8) | |
| Top 30 DRGs | 0·942 (0·002) | 0·995 (0·000) | 0·375 (0·008) | 0·484 (0·009) | 30 | 976 (35·5) | |
| LSTM with clinical measurements | All DRGs | 0·838 (0·017) | 0·946 (0·006) | 0·037 (0·004) | 0·160 (0·012) | 517 | 2747 (100·0) |
AUC and F1 scores on individual DRGs were macro-averaged or micro-averaged within each fold of the experiments, then the results were summarized in mean (standard deviation) over the performances of five models on the hold-out test set. DRG sets refer to stays in the test set used for evaluation that were assigned with either all DRGs (the original test set) or certain subgroups of DRGs (subsets of the test set). The DRG subgroups include most frequent DRGs and DRGs in the three most frequent MDCs, namely diseases and disorders of circulatory system (MDC 05), nervous system (MDC 01), and respiratory system (MDC 04). The number of DRG targets counts the unique DRGs in the DRG category, followed by the number (%) of hospital stays in the test set that were covered in the evaluation. For the model using clinical measurements, we only reported results using all stays as comparison.
DRG diagnostic related group, MS-DRG Medicare severity-DRG, APR-DRG all patient refined-DRG, AUC area under the receiver operating characteristic curve, MDC major diagnosis category.