Skip to main content

Table 3 Predictive accuracy measures by predictive model

From: Building a predictive model of low birth weight in low- and middle-income countries: a prospective cohort study

Measures a

Predictive model

Decision tree

Random forest

Logistic regression

K-nearest neighbors

Linear support vector machine

Area under the curve score

0.64

0.71

0.72

0.58

0.71

Accuracy score

0.62

0.58

0.61

0.83

0.59

Among LBW babies

 Precision

0.21

0.21

0.22

0.20

0.21

 Recall

0.61

0.74

0.72

0.07

0.75

 f1-score

0.31

0.33

0.34

0.10

0.33

Permutation feature importance b – top variables in order of importance

1. maternal weight

2. hypertensive disorders

1. maternal weight

2. clinical site

3. hypertensive disorder

4. antenatal care

5. maternal height

6. antepartum hemorrhage

7. previous livebirth

8. parity

1. clinical site

2. maternal weight

3. antenatal care

4. hypertensive disorder

5. antepartum hemorrhage

6. severe infection during delivery

7. maternal height

1. maternal weight

2. maternal height

3. socio-economic status

4. antenatal care

5. parity

6. previous livebirth

7. maternal education

8. vitamin / calcium supplementation

1. clinical site

2. maternal weight

3. antenatal care

4. hypertensive disorder

5. antepartum hemorrhage

6. maternal education

7. severe infection during delivery

Hyperparameter values c

Balanced class weights, Gini impurity criterion, minimum samples in a leaf = 2, minimum samples required for a split = 4, all others set to default

Balanced class weights, Gini impurity criterion, number of tree = 500, maximum depth = 8, all other set to default

Balanced class weights, L2 ridge regulation penalty, and maximum number of iterations for the solver to converge = 5000

Weights set by distance, leaf size = 20, all other set to default

Linear kernel, balanced class weights, regulation parameter (C) = 100, all other set to default

  1. The binary outcome for the predictive models was LBW. Predictors included maternal age, maternal education, parity, maternal height, maternal weight, socioeconomic status (SES) score, previous livebirth, number of antenatal care visits, use of iron supplementation, use of vitamin or calcium supplementation, hypertensive disorders, severe antepartum hemorrhage, and severe infection during pregnancy. To develop the models, data were split into a training dataset (75% of available data) and test dataset (25% of available data). Hyperparameters were tuned using tenfold grid-search cross validation with scoring = ’roc_auc’. In addition to hyperparameter tuning, we varied the cut point for the probability used to classify an outcome as LBW for the logistic regression model from 0.1 to 0.9. We trained the models on the training data and validated the models on the test data
  2. a Predictive accuracy measures, including area under the curve (AUC) and receiver operator characteristic (ROC) curves, were produced using the test dataset. Precision, recall, and f1 scores were calculated using the classification_report method. Accuracy is (true positives + true negatives) / total. Precision is true positives / (true positives + false positives) and is also known as the Positive Predictive Value. Recall is the sensitivity or true positives / (true positives + false negatives). The f1-score is 2*(precision*recall) / (precision + recall); it is the harmonic mean of precision and recall
  3. b The permutation-based importance was implemented in Scikit-Learn as permutation_importance method. This method randomly shuffles each feature and computes the change in the model’s performance. The features which impact performance the most are the most important ones
  4. c For models which were tuned, the hyperparameter values presented are those in the tuned models
Back to article page

BMC Pregnancy and Childbirth

ISSN: 1471-2393

Contact us