Predicting high health-cost users among people with cardiovascular disease using machine learning and nationwide linked social administrative datasets

Table 2 Predictive performance of identifying high-cost patients with CVD for traditional regression models versus machine learning models

Prediction models	Sensitivity^a	Specificity^b	Positive Predictive Value/ Precision^c	Harmony score (sensitivity & positive predictive value), F1^d	AUC^e
Traditional regression models
All conventional variables (TRM1)^f	4.9%	99.2%	61.2%	9.1%	0.53
As per TRM1 but no ethnicity variables (TRM2)	4.9%	99.2%	62.5%	9.0%	0.53
As per TRM2 but no smoking variables (TRM3)	4.6%	99.2%	61.8%	8.6%	0.53
As per TRM3 but no chronic condition variables	0.0%	100%	Not calculable	Not calculable	0.53
Machine learning models^g
Random forest	37.8%	88.6%	45.2%	41.2%	0.70
KNN	38.0%	85.9%	40.1%	39.0%	0.45
L1-regularised logistic regression	78.9%	83.5%	22.1%	34.5%	0.62
Classification trees	19.5%	98.0%	71.2%	30.6%	0.73

Note: at 20% top high-cost users of all people with CVD
^acorrectly identified people with CVD who are high-cost = recall = true positive rates
^bcorrectly identified people with CVD who are not high-cost = true negative rates
^ctrue positives/false negatives = true positives among people who were predicted positives
^d = 2/(1/recall + 1/precision)
^eArea under the Receiver Operating Characteristics curve
^fage, sex, ethnicity, smoking status, other chronic conditions: diabetes, cancers and traumatic brain injuries
^gall variables as per Table 1

ISSN: 2191-1991