Table 3 Model performance summary for healthcare contacts prediction

From: Machine learning-based predictions of healthcare contacts following emergency hospitalisation using electronic health records

Time point of prediction

Primary outcome metrics

MAE

cMAPE (%)a

BACCb

CKSb

ED Arrival

0.98 [0.98–0.99]

49% [49%–50%]

0.28 [0.28–0.29]

0.34 [0.33–0.35]

Hospital admission

0.93 [0.93–0.94]

47% [46%–47%]

0.32 [0.31–0.33]

0.43 [0.42–0.44]

24 h post-admission

0.86 [0.85–0.87]

42% [41%–43%]

0.32 [0.31–0.32]

0.46 [0.45–0.47]

48 h post-admission

0.82 [0.81–0.83]

36% [35%–37%]

0.32 [0.32–0.33]

0.46 [0.45–0.48]

72 h post-admission

0.80 [0.79–0.81]

34% [33%–34%]

0.32 [0.32–0.33]

0.46 [0.45–0.47]

  1. Values are reported with 95% confidence intervals. Metrics: MAE—mean absolute error, cMAPE—conditional Mean Absolute Percentage Error, BACC—balanced accuracy score, CKS—Cohen’s Kappa score. GMS—geriatric medicine services.
  2. aEstimated by masking 0 values (no linked contacts), due to inherent limitations of MAPE.
  3. bMetrics estimated through quintile-based discretisation of the predicted healthcare contacts.
Back to article page