Data-Driven Insights towards Risk Assessment of Postpartum

Depression

Evdoxia Valavani

, Dimitrios Doudesis

1,2 b

, Ioannis Kourtesis

, Richard F. M. Chin

4,5 d

Donald J. MacIntyre

, Sue Fletcher-Watson

, James P. Boardman

8,9 g

and Athanasios Tsanas

1,10 h

Usher Institute, Medical School, University of Edinburgh, Teviot Place, Edinburgh EH8 9AG, U.K.

BHF Centre for Cardiovascular Sciences, University of Edinburgh, 47 Little France Crescent Edinburgh EH16 4TJ, U.K.

Psychiatric Hospital of Attica Dafni, Athinon Avenue, Athens 12462, Greece

Muir Maxwell Epilepsy Centre, Centre for Clinical Brain Sciences, The University of Edinburgh,

9 Sciennes Road, Edinburgh EH9 1LF, U.K.

Royal Hospital for Sick Children, 9 Sciennes Road, Edinburgh EH9 1LF, U.K.

Division of Psychiatry, Deanery of Clinical Sciences, Royal Edinburgh Hospital, University of Edinburgh,

Morningside Park, Edinburgh EH10 5HF, U.K.

Salvesen Mindroom Research Centre, Kennedy Tower, Royal Edinburgh Hospital, University of Edinburgh,

Morningside Park, Edinburgh EH10 5HF, U.K.

MRC, Centre for Reproductive Health, University of Edinburgh, 47 Little France Crescent, Edinburgh EH16 4TJ, U.K.

Centre for Clinical Brain Sciences, University of Edinburgh, Chancellor’s Building, 49 Little France Crescent,

Edinburgh EH16 4SB, U.K.

Mathematical Institute, University of Oxford, Woodstock Road, Oxford OX2 6GG, U.K.

{d.macintyre, sue.fletcher-watson, james.boardman, athanasios.tsanas}@ed.ac.uk

Keywords: Postpartum Depression, Feature Selection, Random Forests.

Abstract: Postpartum depression is defined as depressive episodes that occur during pregnancy or within 12 months of

parturition. The goal of this study is the exploration of the birth features and maternal traits which affect the

risk of postpartum depression for mothers with preterm neonates. We analysed data from 144 women (63

mothers of term and 81 mothers of preterm infants) who completed the Edinburgh Postnatal Depression Scale

(EPDS) in the postpartum period. We used three feature selection algorithms: ReliefF, Random Forests (RF)

variable importance, and Boruta, in order to select the most predictive feature subsets, which were

subsequently mapped onto the binarized EPDS total score (a threshold of 10 was used to binarize the EPDS

total scores) using RF. We found that positive affectivity (r

=-0.467, p<0.001), and the Apgar score at 5

minutes (r

=-0.430, p<0.001) are the most statistically strongly associated features with the risk of postpartum

depression. We used 10-fold cross-validation with 100 iterations and report out-of-sample balanced accuracy

(median±IQR): 75.0±16.7, sensitivity: 66.7±16.7, specificity: 100±16.7, and F1 score: 0.8±0.2. Collectively,

these findings highlight the potential of using a data-driven process to automate risk prediction using standard

clinical characteristics and motivate the deployment of the developed tool using larger-scale datasets.

https://orcid.org/0000-0003-0625-9454

https://orcid.org/0000-0001-6699-9476

https://orcid.org/0000-0001-9197-4924

https://orcid.org/0000-0002-7256-3027

https://orcid.org/0000-0001-6963-1335

https://orcid.org/0000-0003-2688-1734

https://orcid.org/0000-0003-3904-8960

https://orcid.org/0000-0002-0994-8100

382

Valavani, E., Doudesis, D., Kourtesis, I., Chin, R., MacIntyre, D., Fletcher-Watson, S., Boardman, J. and Tsanas, A.

Data-Driven Insights towards Risk Assessment of Postpartum Depression.

DOI: 10.5220/0009369303820389

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 4: BIOSIGNALS, pages 382-389

ISBN: 978-989-758-398-8; ISSN: 2184-4305

1 INTRODUCTION

Non-psychotic mental disorders are common during

pregnancy and the postpartum period, and have a

negative impact on the mother and her offspring

(Howard et al., 2014). Postpartum depression is

defined by the American Psychiatric Association’s

Diagnostic and Statistical Manual, Fifth Edition

(DSM-5) as a major depressive episode that occurs

either during pregnancy or within 4 weeks of

parturition (APA, 2013). In practice, it is

operationalised as depressive episodes occurring

within 6 to 12 months of delivery (Committee on

Obstetric Practice, 2015). Postnatal non-psychotic

depression affects approximately 3-6% of women

during pregnancy or in the postnatal period (APA,

2013).

Clinical manifestations include despondency,

emotional lability, tearfulness, feelings of guilt,

anhedonia, low energy, loss of appetite, poor sleep,

poor concentration and memory, fatigue, irritability,

as well as feelings of being incapable of looking

after the offspring (Stewart, 2005). It is also

common for some women to report suicidal ideation

(Martini et al., 2019). Many studies have shown that

mothers’ perinatal depressive symptoms can

negatively affect the social, emotional, and

cognitive developmental outcomes of their offspring

(Letourneau et al., 2019).

Risk factors for postpartum depression include

previous history of mental health disorders such as

depression and anxiety, negative feelings towards

the current birth, low self-esteem regarding the

parental role, and positive history of sexual abuse

(Ghaedrahmati et al., 2017). Risky pregnancy,

emergency caesarean section, complications and

long hospitalization also seem to have a negative

impact on mothers’ mental health, whereas

breastfeeding is associated with lower risk of

depression. Furthermore, studies have shown a

positive correlation between young age during

pregnancy, metabolism disorders, reduced

serotonin, oxytocin, oestrogen levels and postnatal

depressed mood (Ghaedrahmati et al., 2017).

Finally, lack of emotional and financial support,

unemployment and low socioeconomic status,

sexual and domestic violence during the prenatal

period, sleeping disorders, as well as habits such as

smoking, low physical activity and unhealthy diet

can potentially increase the risk of postpartum mood

disorders (Ghaedrahmati et al., 2017).

Early detection of postpartum depression is

alarmingly low (Moraes et al., 2017). One standard

widely used clinical metric which is a validated

screening tool for postpartum depression is the

Edinburgh Postnatal Depression Scale (EPDS)

(Cox, Holden and Sagovsky, 1987). It consists of ten

items which assess symptoms of depression such as

anxiety, low energy, sleep disorders, and suicidal

thoughts. The scale shows how the mother has felt

over the preceding week. Total scores of ten or more

indicate possible depression of varying severity. The

EPDS may be used within eight weeks postpartum

and it can also be used as a depression screening tool

during pregnancy (Cox, Holden and Sagovsky,

1987). According to a systematic review of

validation studies of the EPDS conducted by Gibson

et al. 2009, when using a cut-off point of 10, the

sensitivity of the EPDS for detecting postpartum

depression ranged from 59 to 100%, and the

specificity ranged from 44 to 97% (Gibson et al.,

2009).

The aim of this study is to explore the birth

features and maternal traits that affect the risk of

postpartum depression for mothers and build a

statistical learning model in order to predict the risk

of postpartum depression for mothers who have

given birth to preterm neonates.

2 DATA

The study used data that comes from a longitudinal

study conducted at the Royal Infirmary of

Edinburgh which investigates the impact of preterm

birth on long-term outcomes

(https://theirworld.org). A total of 144 women

completed the EPDS during the first month

following childbirth, of whom 63 had term born

neonates, and 81 had preterm neonates. We analysed

data comprising demographic characteristics,

infants’ characteristics (i.e. gestational age,

birthweight, occipitofrontal circumference, Apgar

score), morbidities, and complications that occur

during the neonatal period, socioeconomic status,

and data derived from questionnaires which assess

characteristics of temperament (Adult Temperament

Questionnaire, short form), intelligence (National

Adult Reading Test), and risk of postpartum

depression (Edinburgh Postnatal Depression Scale).

Socioeconomic status of the family was determined

using the Scottish Index of Multiple Deprivation

2016. Table 1 presents the demographic

characteristics of the mothers.

Data-Driven Insights towards Risk Assessment of Postpartum Depression

383

Table 1: Demographic characteristics of the mothers who

completed the EPDS in the postpartum period.

Characteristic

Mothers with

term born

neonates

(n=63)

Mothers with

preterm

neonates

(n=81)

Age (years) 34 ± 7 31 ± 7

Ethnicity

Any White 54 (91) 73 (93)

Asian 2 (4) 4 (5)

Black 0 (0) 1 (1)

Mixed 2 (3) 1 (1)

Other 1 (2) 0 (0)

Education

None 0 (0) 2 (3)

1-4 GCSE 1 (2) 4 (6)

≥5 GCSE 2 (3) 5 (7)

A Levels/Highers 2 (3) 9 (12)

College 4 (7) 17 (24)

University 27 (47) 22 (31)

Postgraduate degree 22 (38) 13 (18)

SIMD Quintile

1 3 (5) 10 (13)

2 8 (13) 22 (28)

3 11 (17) 12 (15)

4 17 (27) 18 (23)

5 24 (38) 16 (21)

Full IQ 122.74 ± 4.96 121.10 ± 7.43

BMI 24.7 ± 6.6 26.2 ± 7.28

Gestation (weeks) 39.56 ± 1.42 29.84 ± 2.72

Variables are presented in the form median ± IQR or

number (%). Education refers to the highest educational

level attained. GCSE stands for General Certificate of

Secondary Education. SIMD stands for Scottish Index of

Multiple Deprivation. IQ stands for Intelligence Quotient.

BMI stands for Body Mass Index.

3 METHODS

We used the standard clinical cut-off of 10 (Cox,

Holden and Sagovsky, 1987) to binarize the EPDS

total scores; scores lower than 10 were classified as

“low risk for postpartum depression”, and scores

equal to or higher than 10 were classified as “high risk

for postpartum depression”. We aimed to build a

statistical model in order to estimate whether scores

would be above or below the clinical threshold (i.e.

this is a binary classification task).

3.1 Preliminary Statistical Analysis

We used the Mann – Whitney rank sum test (Mann

and Whitney, 1947) to test whether the difference

between the medians of the EPDS total score of

mothers with term born neonates and mothers with

preterm neonates is zero. The level of significance

associated with the null hypothesis was set at

α = 0.05.

In order to quantify the strength of the association

between two variables, we computed the Spearman’s

rank correlation coefficient which is effective in

quantifying general monotonic relationships

(Schober, Boer and Schwarte, 2018). There is no

universal guideline to determine when a bivariate

relationship is statistically strong; it depends on the

application (Cohen et al., 2002). In this study, we

consider that an absolute value of a correlation

coefficient > 0.3 corresponds to a statistically strong

association, in accordance to similar studies in

clinical contexts (Meyer et al., 2001; Tsanas, Little

and McSharry, 2013).

3.2 Feature Selection

A problem that often arises when we analyse high

dimensional data is the curse of dimensionality

(Bellman, 1966); given that there is only a finite

number of available samples, as the dimensions of the

feature space increase, it is difficult to adequately

populate the feature space, with detrimental effects in

the performance of the learners (Hastie, Tibshirani

and Friedman, 2009). The number of data samples

required, often grows exponentially with the number

of features. Prediction performance can often be

improved by reducing the number of dimensions in

the feature space, a process that is called

dimensionality reduction. Dimensionality reduction

techniques can be divided into two main categories:

feature transformation and feature selection (Guyon

and Elisseeff, 2003).

In this study, we have compared three efficient

feature selection algorithms: (a) ReliefF (Kononenko,

1994), (b) Random Forests (RF) variable importance

(Hastie, Tibshirani and Friedman, 2009; James et al.,

2013), and (c) Boruta (Kursa and Rudnicki, 2010).

ReliefF is a feature-weighting algorithm; it

assigns a ‘weight’ value to all features of a dataset

based on how well their values distinguish between

the data samples that are near to each other and thus,

how useful they are in predicting the response

variable. There are various feature evaluation

measures that ReliefF can use. In this study, we have

explored ReliefF expRank where k nearest instances

have weight exponentially decreasing with increasing

rank.

When using RF (Breiman, 2001), it is possible to

measure the importance of each feature in predicting

SERPICO 2020 - Special Session on Mining Self-reported Outcome Measures, Clinical Assessments, and Non-invasive Sensor Data

Towards Facilitating Diagnosis, Longitudinal Monitoring, and Treatment

384

the response variable. At each split in each tree, the

improvement in the split – criterion is the importance

measure attributed to the splitting variable and is

accumulated over all the trees in the forest for each

feature. In the context of classification models, we

can measure the total amount that the Gini index is

decreased by splits over a given feature, averaged

over all trees. A large value indicates an important

feature (Hastie, Tibshirani and Friedman, 2009;

James et al., 2013).

The Boruta algorithm is a wrapper feature

selection technique built around the random forest

learner that measures the importance of each feature

by dividing the average loss of accuracy among all

trees by the standard deviation of the accuracy loss

(Kursa and Rudnicki, 2010). In other words, it uses Z

score as the importance measure.

In this study, the feature subsets were selected

using a Cross-Validation (CV) approach, using only

the training dataset in each CV iteration. We repeated

the CV process a total of 100 times. The feature

selection algorithms described above aim at ranking

the features of a given dataset based on their

contribution towards prediction of the response

variable. In order to select the final feature subset for

each feature selection algorithm, we followed the

process described by Tsanas et al. in 2012 (Tsanas et

al., 2012). In a nutshell, for a given feature selection

algorithm, when using CV, at the end of each

iteration, we obtain a vector of the ordered sequence

of the indices of the features, where the first feature is

considered to be the most important one, and the last

feature corresponds to the least important one. We

store these vectors in a matrix of n x p size, where n

corresponds to the number of iterations and p

corresponds to the number of features of a given

dataset. This way, in each of the rows of the matrix

we have stored the feature subset selected at the end

of each iteration. Subsequently, we need to identify

the feature index which has most frequently been

ranked as first across all iterations, then we need to

identify which feature appears most frequently as

second or third and so on. In case a feature index has

already been included in the final subset and is later

found again as most frequent, we need to select the

second most frequent and so on. Ties are resolved by

including the lowest index number. In the end, we

obtain a vector which consists of feature indices

ordered from the most important to the least

important one, for a given feature selection algorithm.

These features are then used to train the learner one

by one. We choose to keep the most parsimonious

model that gives the best overall results in terms of

accuracy, sensitivity, specificity, and F1 score.

3.3 Mapping Selected Features to the

Response

Our ultimate goal is to estimate the function f which

relates X and y, that is f(X) = y, where y is the

response variable and X is a matrix comprising p

features X

, X

, …, X

. That is, we need a binary

classifier that will use the selected features to predict

the risk of postpartum depression.

Here, we used RF (Breiman, 2001) which is a

widely-used statistical machine learning algorithm.

RF involve fitting decision trees on bootstrapped

samples of the original training data and then

combining all individual trees (we used 500 trees) to

create a single powerful predictive model. When

building these decision trees, each time a split in a tree

is considered, a random sample of m features is

chosen as split candidates from the full set of p

features. The split is allowed to use only one of those

m predictors. Following the suggestion of Breiman

(Breiman, 2001), we set m to be equal to the square

root of the number of features in the training data.

3.4 Classifier Validation

We validated our model using CV, which provides an

estimate of the performance of the model on new

“unseen” data, provided that the new data comes from

the same joint distribution as the data that was used to

train the model. In this study, we used 10-fold cross-

validation and repeated the process 10 times,

resulting in 100 iterations.

4 RESULTS

4.1 Preliminary Statistical Analysis

Regarding mothers who had term born neonates (n =

63), the total scores of the EPDS ranged from 0 to 15,

with a median score of 6. Among them, 16 mothers

had a total score of 10 or above, indicating high risk

of depression, while the rest scored lower than 10.

With regards to the mothers who had preterm

neonates (n = 81), the total scores of the EPDS ranged

from 0 to 23, with a median score of 7. Among them,

27 mothers had a total score of 10 or above, indicating

high risk of depression, while the rest of the mothers

scored lower than 10.

Figure 1 shows a histogram of the total scores of

the EPDS for mothers with term born neonates and

mothers with preterm neonates. The dashed lines

represent the median scores. The p-value of the Mann

Data-Driven Insights towards Risk Assessment of Postpartum Depression

385

– Whitney rank sum test is 0.459. At the 0.05

significance level, there is not enough evidence to

reject the null hypothesis that there is no statistically

significant difference in the medians of the EPDS

total score between mothers who had a term born

neonate and those who gave birth to preterm

neonates.

Figure 1: EPDS total score histogram plot. The dashed line

represents the median scores.

Table 2: Correlation analysis for the EPDS total score of the

mothers who had preterm neonates (n = 81).

Feature

Statistical

association

(Spearman

correlation

coefficient)

Statistical

significance

(p-value)

Positive affect -0.467 <0.001

Apgar 5min -0.430 <0.001

Fear 0.399 <0.01

Extraversion -0.347 <0.01

Attentional control -0.339 <0.01

Days of intubation 0.330 <0.01

Days of supplemental O

0.323 <0.01

Gestation -0.318 <0.01

Previous live births 0.309 <0.01

Neutral perceptual

sensitivity

-0.307 <0.05

Only the ten most strongly associated features with the

EPDS total score are presented in the Table for brevity.

Positive affect, fear, extraversion, attentional control, and

neutral perceptual sensitivity are characteristics of

temperament derived from the Adult Temperament

Questionnaire, short form, and assessed during the first

month following delivery.

Table 2 presents the ten features (out of 102) most

strongly associated with the total score of the EPDS

for mothers with preterm infants. The features are

sorted according to the absolute value of the

correlation coefficient. Positive affect, fear,

extraversion, attentional control, previous live births,

and neutral perceptual sensitivity are all maternal

features. Apgar score at 5 minutes, total number of

days of intubation, total number of days of

supplemental oxygen, and gestation refer to infant

features. It is interesting to note that all features

exhibit statistically significant (p<0.05) correlation.

These findings give some initial confidence that the

binary classification task of this study may lead to

accurate results.

4.2 Classification Stage: Mapping

Features to the Thresholded EPDS

Total Score for Mothers with

Preterm Neonates

Figures 2, 3 and 4 illustrate the out of sample

performance of the RF as a function of the number of

features selected by the different feature selection

algorithms. We found that the feature size giving the

best results in terms of accuracy, sensitivity,

specificity, and F1 score is 13 using ReliefF expRank.

Therefore, our subsequent results use the first 13

features selected by each feature selection algorithm.

Table 4 presents the selected feature subsets for each

feature selection algorithm. We remark that reducing

Figure 2: Comparison of out of sample median balanced

accuracy results with IQR of the random forests classifier

using the features selected by each of the five feature

selection algorithms. These results are computed using 10-

fold CV with 100 iterations. For clarity, only the first 30

steps are presented.

SERPICO 2020 - Special Session on Mining Self-reported Outcome Measures, Clinical Assessments, and Non-invasive Sensor Data

Towards Facilitating Diagnosis, Longitudinal Monitoring, and Treatment

386

the original 102 dimensions of the feature space leads

to an improvement in out-of-sample performance

accuracy with RF. Table 3 shows a confusion matrix

of the performance of the RF when mapping the

features selected by ReliefF expRank (i.e. previous

live births, bronchopulmonary dysplasia, and

extraversion) to the response.

Figure 3: Comparison of out of sample median sensitivity

results with IQR of the random forests classifier using the

features selected by each of the five feature selection

algorithms. These results are computed using 10-fold CV

with 100 iterations. For clarity, only the first 30 steps are

presented.

Figure 4: Comparison of out of sample median specificity

results with IQR of the random forests classifier using the

features selected by each of the five feature selection

algorithms. These results are computed using 10-fold CV

with 100 iterations. For clarity, only the first 30 steps are

presented.

Table 3: Confusion matrix summarizing the out-of-sample

findings using 10-fold CV.

Reference

High risk of

depression

Low risk of

depression

Prediction

High risk of

depression

165 47

Low risk of

depression

105 493

Total number of entries is 810, which is derived from 81

samples x 10 iterations.

Table 4: Selected feature subsets and out-of-sample

classification performance.

ReliefF

expRanK

RF variable

importance

Boruta

Previous live

births

Extraversion Extraversion

BPD Positive affect Positive affect

Extraversion Fear Attentional

control

Maternal

anxiety

Gestation Fear

Days of

intubation

Attentional

control

Gestation

Apgar score

at 5 min

Inhibitory

control

Apgar score at

5 min

Days of

supplemental

Effortful

control

Inhibitory

control

Effortful

control

Apgar score at

5 min

Effortful

control

Inhibitory

control

Affective

perceptual

sensitivity

Affective

perceptual

sensitivity

Positive

affect

Days of supple-

mental O

Days of supple-

mental O

Feeding at

discharge

from NICU

Previous live

births

Days of

intubation

ROP Negative affect Negative affect

Maternal

asthma

Total days in

the NICU

Total days in

the NICU

Accuracy

Specificity

Sensitivity

F1 score

75.0±16.7

100±16.7

66.7±16.7

0.8±0.2

83.3±18.3

100±16.7

66.7±50.0

0.8±0.2

75.0±23.3

100±20.0

66.7±66.7

0.8±0.2

The selected features from each feature selection algorithm

are sorted in descending order from the most important to the

least important one. The last row presents the performance of

the random forests classifier for each feature subset. The

results are given in the form median±IQR. BPD stands for

Bronchopulmonary Dysplasia. NICU stands for Newborn

Intensive Care Unit. ROP stands for Retinopathy of

Prematurity.

Data-Driven Insights towards Risk Assessment of Postpartum Depression

387

5 DISCUSSION

In this study, our goal has been to build a

parsimonious model to determine the features that

could potentially predict the risk of postpartum

depression for mothers with preterm infants. To this

end, we compared three different feature selection

algorithms: (a) ReliefF, (b) Boruta, and (c) RF

variable importance in order to derive a small feature

subset which would be more informative than the

original feature set. Subsequently, in order to map the

selected features to the response we used RF.

We analysed data from 144 women who

completed the EPDS during the postpartum period, of

whom 63 have had term born neonates, and 81 have

had preterm neonates. For mothers who gave birth to

preterm neonates, correlation analysis revealed

statistically strong correlations. At this point, we need

to emphasize that interpretation of the results is

tentative due to the small sample size. We found that

the risk of postpartum depression is affected both by

mothers’ temperament, and perinatal factors.

Specifically, the two most strongly associated

features with risk of postpartum depression for

mothers with preterm infants are positive affect (r

0.467, p<0.001), and the Apgar score at 5 minutes

after birth (r

=-0.430, p<0.001). These two features

have a negative association with postpartum

depression, meaning that mothers with high positive

affectivity (i.e. enthusiastic, energetic, confident)

have a lower risk of developing postpartum

depression. A lower Apgar score at 5 minutes after

birth increases the risk of postpartum depression.

Here, the Apgar score might be a proxy for

subsequent events, such as neonatal morbidities or

days spent in the NICU. Fear, another characteristic

of adult temperament assessed by the Adult

Temperament Questionnaire short form, has a strong

positive association with risk of postpartum

depression (r

=0.399, p<0.01). On the other hand,

extraversion (i.e. the tendency of primarily obtaining

gratification from outside oneself) (r

=-0.347,

p<0.01), attentional control (i.e. capacity to focus and

shift attention when desired) (r

=-0.339, p<0.01), and

neutral sensitivity (i.e. detection of low intensity

stimuli) (r

=-0.307, p<0.05) have a strong negative

association with postpartum depression. Moreover,

risk of postpartum depression increases with

decreasing gestational age (r

=-0.318, p<0.01). The

total days of intubation (r

=0.330, p<0.01), as well as

the total days of supplemental O

=0.323, p<0.01)

that a preterm infant requires increase the risk of

postpartum depression. These two exposures are

proxies of respiratory co-morbidities of preterm birth,

which may cause distress to mothers. Finally, a past

history of live births is positively associated with

postpartum depression (r

=0.309, p<0.01).

We found that the feature size giving the best

results in terms of accuracy, sensitivity, specificity,

and F1 score is 13 using the feature selection

algorithm ReliefF expRank. The feature subset

selected by the ReliefF expRank comprised past

history of live births, bronchopulmonary dysplasia,

extraversion, anxiety, total days of intubation and

supplemental oxygen, Apgar score at 5 minutes after

birth, effortful control, inhibitory control, positive

affect, the infant’s feeding at discharge, retinopathy

of prematurity, and maternal asthma. We can

conclude that extraversion, which is a characteristic

of adult temperament, positive affect, as well as

effortful control and inhibitory control are negatively

correlated with the risk of postpartum depression. On

the other hand, past history of anxiety and chronic

diseases such as asthma may increase the risk of

postpartum depression. Bronchopulmonary

dysplasia, and retinopathy of prematurity are diseases

that can result in significant morbidity for preterm

neonates, causing distress to mothers. Furthermore, a

lower Apgar score at 5 minutes after birth, and a

higher number of days requiring intubation and

supplemental oxygen increase the risk of postpartum

depression; mothers who breastfeed their babies have

a lower risk of depression compared to mothers

whose babies drink infant formula. Finally, mothers

with existing childcare responsibilities find it more

challenging to cope after having a preterm infant

compared with those who do not have pre-existing

childcare duties. The performance of the RF resulted

in out-of-sample balanced accuracy (median±IQR):

75.0±16.7, specificity: 100±16.7, and sensitivity:

66.7±16.7, and F1 score: 0.8±0.2.

Although we found some statistically strong

correlations which gave some initial confidence that

the statistical learning model would have a good

chance of success, the performance of the RF was

poor when taking into account the class imbalance in

the investigated problem. One possible reason for the

poor model performance is lack of statistical power:

the sample size is not sufficiently large to effectively

train the models. Another reason for poor model

performance might be the fact that the given variables

are insufficient to predict the response accurately. In

addition, we need to bear in mind that a considerable

number of features has been derived from

questionnaires completed by the individuals

themselves and therefore subjectivity in data

collection may be another issue affecting the results.

SERPICO 2020 - Special Session on Mining Self-reported Outcome Measures, Clinical Assessments, and Non-invasive Sensor Data

Towards Facilitating Diagnosis, Longitudinal Monitoring, and Treatment

388

We remark there are many contemporary

advanced classifiers that can be used in order to map

the features to the response, such as RF and Support

Vector Machines (SVM). Here, in order to map the

selected features to the response we only used RF and

have not explored competing approaches (e.g. SVM)

to optimize the classification performance further.

We envisage that a larger sample size may lead to

a better and more generalizable model that may be

more accurate in correctly assessing postpartum

depression risk. Future work could potentially

incorporate additional statistical machine learning

algorithms which may improve prediction accuracy.

REFERENCES

APA (2013) DSM-V-The Diagnostic and Statistical

Manual of Mental Disorders.

Bellman, R. (1966) ‘Dynamic programming.’, Science

(New York, N.Y.). American Association for the

Advancement of Science, 153(3731), pp. 34–7. doi:

10.1126/science.153.3731.34.

Breiman, L. (2001) ‘Random Forests’, Machine Learning.

Kluwer Academic Publishers, 45(1), pp. 5–32. doi:

10.1023/A:1010933404324.

Cohen, J. et al. (2002) Applied Multiple Regression/

Correlation Analysis for the Behavioral Sciences Third

Edition. Available at: https://www.gbv.de/dms/

ilmenau/toc/348809573.PDF (Accessed: 28 June

2019).

Committee on Obstetric Practice (2015) ‘The American

College of Obstetricians and Gynecologists Committee

Opinion no. 630. Screening for perinatal depression.’,

Obstetrics and gynecology, 125(5), pp. 1268–71. doi:

10.1097/01.AOG.0000465192.34779.dc.

Cox, J. L., Holden, J. M. and Sagovsky, R. (1987)

‘Detection of Postnatal Depression’, British Journal of

Psychiatry. Cambridge University Press, 150(06), pp.

782–786. doi: 10.1192/bjp.150.6.782.

Ghaedrahmati, M. et al. (2017) ‘Postpartum depression risk

factors: A narrative review.’, Journal of education and

health promotion. Wolters Kluwer -- Medknow

Publications, 6, p. 60. doi: 10.4103/jehp.jehp_9_16.

Gibson, J. et al. (2009) ‘A systematic review of studies

validating the Edinburgh Postnatal Depression Scale in

antepartum and postpartum women’, Acta Psychiatrica

Scandinavica. John Wiley & Sons, Ltd (10.1111),

119(5), pp. 350–364. doi: 10.1111/j.1600-

0447.2009.01363.x.

Guyon, I. and Elisseeff, A. (2003) An Introduction to

Variable and Feature Selection, Journal of Machine

Learning Research. Available at:

http://jmlr.csail.mit.edu/papers/volume3/guyon03a/gu

yon03a.pdf (Accessed: 29 June 2019).

Hastie, T., Tibshirani, R. and Friedman, J. (2009) The

Elements of Statistical Learning Data Mining,

Inference, and Prediction. 2nd edn. Springer New

York.

Howard, L. M. et al. (2014) ‘Non-psychotic mental

disorders in the perinatal period.’, Lancet (London,

England), 384(9956), pp. 1775–88. doi: 10.1016/S01

40-6736(14)61276-9.

James, G. et al. (2013) An Introduction to Statistical

Learning with Applications in R.

Kononenko, I. (1994) ‘Estimating attributes: Analysis and

extensions of RELIEF’, in. Springer, Berlin, Heidelberg,

pp. 171–182. doi: 10.1007/3-540-57868-4_57.

Kursa, M. B. and Rudnicki, W. R. (2010) ‘Feature Selection

with the Boruta Package’, Journal of Statistical

Software, 36(11), pp. 1–13. doi: 10.18637/jss.v036.i11.

Letourneau, N.

et al. (2019) ‘Maternal and paternal

perinatal depressive symptoms associate with 2- and 3-

year-old children’s behaviour: findings from the

APrON longitudinal study’, BMC Pediatrics, 19(1), p.

435. doi: 10.1186/s12887-019-1775-1.

Mann, H. B. and Whitney, D. R. (1947) ‘On a Test of

Whether one of Two Random Variables is

Stochastically Larger than the Other’, The Annals of

Mathematical Statistics. Institute of Mathematical

Statistics, 18(1), pp. 50–60. doi: 10.1214/aoms/117773

0491.

Martini, J. et al. (2019) ‘Predictors and outcomes of suicidal

ideation during peripartum period.’, Journal of affective

disorders, 257, pp. 518–526. doi: 10.1016/j.jad.2019.

07.040.

Meyer, G. J. et al. (2001) ‘Psychological testing and

psychological assessment. A review of evidence and

issues.’, The American psychologist, 56(2), pp. 128–65.

Available at: http://www.ncbi.nlm.nih.gov/pubmed/

11279806 (Accessed: 7 July 2019).

Moraes, G. P. de A. et al. (2017) ‘Screening and diagnosing

postpartum depression: when and how?’, Trends in

Psychiatry and Psychotherapy. Associação de

Psiquiatria do Rio Grande do Sul, 39(1), pp. 54–61. doi:

10.1590/2237-6089-2016-0034.

Schober, P., Boer, C. and Schwarte, L. A. (2018)

‘Correlation Coefficients’, Anesthesia & Analgesia,

126(5), pp. 1763–1768. doi: 10.1213/ANE.00000000

00002864.

Stewart, D. (2005) ‘Depression during pregnancy.’,

Canadian family physician Medecin de famille

canadien. College of Family Physicians of Canada,

51(8), pp. 1061–7. Available at: http://www.ncbi.nlm.

nih.gov/pubmed/16121822 (Accessed: 2 December

2019).

Tsanas, A. et al. (2012) ‘Novel Speech Signal Processing

Algorithms for High-Accuracy Classification of

Parkinson’s Disease’, IEEE Transactions on

Biomedical Engineering, 59(5), pp. 1264–1271. doi:

10.1109/TBME.2012.2183367.

Tsanas, A., Little, M. A. and McSharry, P. E. (2013) ‘A

methodology for the analysis of medical data’, in

Handbook of Systems and Complexity in Health.

Springer, pp. 113–125.

Data-Driven Insights towards Risk Assessment of Postpartum Depression

389