Prediction of QT Prolongation in Advanced Breast Cancer Patients

Using Survival Modelling Algorithms

Asmir Vodenčarević

, Julia Kreuzeder

, Achim Wöckel

and Peter A. Fasching

Innovative Medicines, Novartis Pharma GmbH, Nuremberg, Germany

Department of Obstetrics and Gynecology, University Hospital Würzburg, Würzburg, Germany

Department of Obstetrics and Gynecology, University Hospital Erlangen, Erlangen, Germany

Keywords: QT Prolongation, Survival Modelling, Advanced Breast Cancer, Clinical Trial Data, Real-World Data.

Abstract: Advanced breast cancer includes locally advanced disease and metastatic breast cancer with distant metastasis

in other organs like lung, liver, brain and bone. While it cannot be cured, its progression can be controlled by

modern treatments including targeted therapies. However, these therapies as well as certain risk factors like

advanced age can facilitate toxicities such as prolongation of the time interval between the start of the Q wave

and the end of the T wave in patient’s electrocardiogram. This could lead to serious life-threatening issues

like cardiac arrhythmia. In this paper we addressed the issue of individual, patient-level prediction of QT

prolongation in advanced breast cancer patients treated with the CDK4/6-inhibitor ribociclib. By formulating

the prediction task as a survival analysis problem, we were able to apply five conventional statistical and

machine learning survival modelling algorithms to both clinical trial and real-world data in order to train and

externally validate prediction models. Cox proportional hazards model regularized by elastic net reached

external, cross-study validation performance (c-index based on inverse probability of censoring weights) of

0.63 on the real-world data and 0.71 on the clinical trial data. The most important predictive factors included

baseline electrocardiogram features and patient quality of life.

1 INTRODUCTION

Breast cancer is the most frequent female cancer

worldwide (Arnold et al., 2022). In 2020, there have

been more than 2.3 million new cases and 685,000

deaths recorded, with the tendency to reach 3 million

new cases and 1 million deaths in 2040 (Arnold et al.,

2022). If not diagnosed and treated early, it can spread

to other organs like liver, lungs, brain and even bones.

Although such advanced (also called metastatic)

breast cancer is considered incurable, its progression

and symptoms can be kept under control by

treatments such as chemotherapy, radiotherapy,

immunotherapy, hormone and targeted therapy. An

important type of targeted therapy are Cyclin-

Dependent Kinase 4 and 6 (CDK4/6) inhibitors.

These relatively new drugs block the activity of

CDK4/6 kinases, which are crucial for growth and

division of cancer cells. In this way, they can improve

https://orcid.org/0000-0002-1120-7547

https://orcid.org/0000-0002-6767-9666

https://orcid.org/0000-0003-1709-1079

Figure 1: Illustration of QT prolongation in patient’s

electrocardiogram (Brody, 2016).

survival of patients as well as their quality of life

considerably (Lu Y.S. et al., 2022). However, some

therapies are associated with potentially serious

164

VodenÄ areviÄ

G, A., Kreuzeder, J., WÃ˝uckel, A. and Fasching, P.

Prediction of QT Prolongation in Advanced Breast Cancer Patients Using Survival Modelling Algorithms.

DOI: 10.5220/0012130900003541

In Proceedings of the 12th International Conference on Data Science, Technology and Applications (DATA 2023), pages 164-172

ISBN: 978-989-758-664-4; ISSN: 2184-285X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

toxicities including prolongation of the time interval

between the start of the Q wave and the end of the T

wave in patient’s electrocardiogram (ECG) (Ward et

al., 2019), as illustrated in Figure 1 (Brody, 2016). An

extended QT interval can lead to cardiac arrhythmia

and in some cases to sudden cardiac death. QT

prolongation is part of the toxicity assessment during

every new medication approval process. Many drugs

associated with a QT prolongation have been

approved. During their application, QT intervals need

to be closely monitored in treated patients. Clinically,

it would be helpful to identify patients who have a

higher or lower risk for a QT prolongation to possibly

adapt the monitoring according to the risk. Individual

risk assessments are based on well-known risk factors

like age or history of cardiovascular diseases. To the

best of our knowledge, there are currently no

published survival modelling approaches to

individual prediction of QT prolongation in advanced

breast cancer with or without treatment with CDK4/6

inhibitors.

The contribution of this paper is three-fold. First,

we present the results of our feasibility study on

predicting QT prolongation in individual advanced

breast cancer patients treated with one of the

prominent CDK4/6 inhibitors ribociclib. Our target

group are patients with the most prevalent subtype of

advanced breast cancer, namely hormone receptor‒

positive / human epidermal growth factor receptor 2‒

negative (HR+/HER2-) advanced breast cancer.

Since our data is only partially observable (outcomes

available only in the course of the clinical studies),

we formulated the QT prolongation prediction task as

a survival analysis problem, which we addressed with

survival modelling algorithms. Second, several linear

and non-linear algorithms are evaluated and

compared. Third, as we had access to both smaller,

high-quality clinical trial data and larger, lower-

quality real-world data, we performed both internal

(within study), nested cross-validation and external,

cross-study validation, training models in one and

validating in another study. This enabled gaining

valuable, potentially generalizable insights in the

utility of both data sources for training statistical and

machine learning survival models to predict clinical

events.

2 RELATED WORK

Survival modelling algorithms have been already

applied to different medical prediction tasks (Spooner

et al., 2020; Qiu et al., 2020) including prediction of

breast cancer survival (Moncada-Torres et al., 2021).

However, there haven’t been many studies in general

aiming at assessing the risk of QT prolongation on

individual, patient-level, especially those treated with

CDK4/6 inhibitors. A retrospective study of large

healthcare claims data (Ward et al., 2019) analysed

risk factors for QT prolongation in HR+/HER2-

metastatic breast cancer patients. These general risk

factors include advanced age, congenital long QT

syndrome, cardiovascular disease, electrolyte

abnormalities and concomitant medication. The Heart

Failure Association of the European Society of

Cardiology jointly with the International Cardio-

Oncology Society has provided tools for baseline

cardiovascular risk assessment in patients scheduled

to receive cardiotoxic cancer drugs (Lyon et al.,

2020). Risk stratification into very high, high and

medium risk based on several patient baseline

characteristics has been proposed, however not for

the CDK4/6 class of drugs.

In (Tisdale et al., 2013) a relatively accurate

statistical model (c-statistic 0.83, sensitivity /

specificity 0.74 / 0.77) for quantification of the QT

prolongation risk based on easily obtainable clinical

variables have been proposed. The model was

developed for and applicable to hospitalized patients

only. A related QT prolongation alert system was

developed and implemented at Mayo Clinic, aiming

at identification of patients under high risk of

mortality (Haugaa et al., 2013). This rule-based

system was derived from the expert knowledge both

for paediatric and adult patients and represented as a

decision tree. A more comprehensive list of risk

factors for QT prolongation (corrected for the heart

rate) was included into the RISQ-PATH score

(Vandael et al., 2017), which was validated in the

Nexus hospital network in Belgium demonstrating

sensitivity of 0.87 and specificity of 0.46 (Vandael et

al., 2018). In (Fasching et al., 2022) the problem of

predicting QT prolongation was treated as a binary

classification task. The same data was used as in our

work and the LASSO method was applied. In one

dataset (RIBECCA study, Decker et al., 2021), the

area under the receiver operating characteristic curve

(AUROC) measured in cross-validation reached 0.67

(weighted AUROC 0.77). However, no predictive

signal was observed in the validation dataset

(AUROC 0.49, weighted AUROC 0.49 in RIBANNA

study, Lüftner et al., 2022). While accurate individual

prediction of QT prolongation is difficult,

understanding its underlaying mechanism remains

even more challenging and might require further

molecular genetic studies (Roden et al., 2016). This

hypothesis is underlined in (Schwartz et al., 2016) by

linking drug-induced and congenital QT

Prediction of QT Prolongation in Advanced Breast Cancer Patients Using Survival Modelling Algorithms

165

prolongation, which could be explained by the

growing genetic evidence in the future.

3 DATA SELECTION AND

PREPARATION

3.1 Study Data

In this work we used anonymized data from two

studies: RIBECCA (Decker et al., 2021) clinical trial

and RIBANNA (Lüftner et al., 2022) non-

interventional study (real-world data). RIBECCA

was a national, multicentre single-arm, open-label

phase 3b clinical trial investigating the efficacy and

safety of treatment with ribociclib (a CDK4/6

inhibitor) plus letrozole in patients with HR+/HER2-

advanced (recurrent or metastatic) breast cancer.

RIBANNA is a still ongoing non-interventional study

evaluating the real-world efficacy and safety of first-

line ribociclib in combination with aromatase

inhibitor/fulvestrant, endocrine monotherapy or

chemotherapy. Description of the original data is

given in the references for these studies.

3.2 Data Selection

3.2.1 Patient Selection

This analysis included patients with available data at

baseline, i.e., at the time point prior to treatment start.

A total of 584 patients (including screening failures)

from RIBECCA and 2316 from RIBANNA were

considered for the analysis. Patients were filtered in

the following hierarchical order: at first, patients who

received at least one dose of study medication are

selected, resulting in 502 and 2211 patients in

RIBECCA and RIBANNA, respectively. Two

patients with non-positive PR interval in ECG were

removed from the RIBECCA data, leaving 500

patients in the final RIBECCA cohort. In the next

step, RIBANNA patients who were not treated with

ribociclib were excluded, leaving 1858 patients in the

analysis. Since RIBANNA contains real-world data

with accordingly lower quality (due to the real-world

treatment and less intense data monitoring as

compared to clinical trial data), we carefully checked

it for any unusual values. One patient with zero blood

pressure (both systolic and diastolic), five patients

with non-positive RR, PR or QRS intervals in ECG

and 12 patients with negative number of days since

primary diagnosis were excluded, resulting in 1840

RIBANNA patients.

3.2.2 Variable Selection

The anonymized RIBECCA and RIBANNA data

included about 420 variables, out of which the

majority are not relevant for our modelling task, e.g.

many absolute dates and placeholders for safety and

tumour control variables. Based on the domain

knowledge, 72 potentially relevant variables were

selected, which were recorded in both studies. This

criterion was a prerequisite for performing external,

cross-study validation. These variables (all recorded

at baseline) served as input data to prediction models,

and they are grouped as follows:

 Demographic characteristics including age and

body-mass index

 Vital signs including ECG features (like PR,

QT and QRS interval), systolic and diastolic

blood pressure, heart rate

 Diagnosis and cancer severity features like

days since primary diagnosis, histological

grade, metastasis location

 Medical history including vomiting,

pneumonia, fatigue

 Prior therapy including most recent prior

therapy, surgery, radiotherapy

 Hormone receptor status

 Eastern Cooperative Oncology Group (ECOG)

patient’s performance status scale and patient

reported outcomes including different EORTC

(European Organisation for Research and

Treatment of Cancer) quality of life

questionnaires

The target variable was QT prolongation. It was

recorded in both studies as a binary event indicator

(QT prolongation has happened or not) together with

the event absolute date. Rather than trying to predict

the target at a single time point or within a specified

time horizon, we formulated the prediction task as a

survival analysis problem. As its name says, survival

analysis traditionally aims at predicting the time to

death and it originates from clinical research. The

target is typically censored, meaning that it is only

observed within an observation period. In the context

of clinical studies, a clinical event can be observed

typically only during study and it either happens or

not. It remains unknown if and when the event has

happened after the study has ended or the patient has

dropped out (discontinued from the study for

whatever reason). We translated the QT prolongation

prediction problem into the survival analysis problem

by (1) computing the time to QT prolongation from

the event date and the baseline date for patients who

experienced it, and (2) computing the time of

censoring for patients who didn’t experience it. In the

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

166

implementation, the target variable was a structured

array of (event, event_time) pairs, where event is a

binary QT prolongation indicator and event_time is a

time of event if QT prolongation has happened or time

of last contact with the patient if it didn’t. As common

for survival analysis problems, the target was

imbalanced. QT prolongation was recorded in 37

(7.4%) RIBECCA patients and 61 (3.3%) RIBANNA

patients with median observation times of 42 and 27

days, respectively. Corresponding Kaplan-Meier

curves, which illustrate the estimated event-free

probability as a function of time, are given in Figures

2 and 3.

Figure 2: Kaplan-Meier curve for QT prolongation in

RIBECCA.

Figure 3: Kaplan-Meier curve for QT prolongation in

RIBANNA.

3.3 Data Preparation for Modelling

All data preparation steps described in this section

were performed in an unsupervised manner, i.e. the

target variable was not considered. After patient and

initial variable selection based on domain knowledge

was performed, the proportion of missing values was

checked. In total, 4.6% and 24.5% of values in the

baseline, input data were missing in RIBECCA and

RIBANNA, respectively, confirming the expected

considerably higher completeness of clinical trial data

comparing to real-world data. Variables containing

more than 50% of missing values in either study were

removed from both studies. This affected only four

variables. In the next step, it was checked for highly

correlated, redundant numerical variables using

Pearson correlation coefficient. Absolute value of

correlation coefficient higher than 0.8 was observed

only between body-mass index and patient weight. As

in (Decker et al., 2021) both weight and height were

removed and body-mass index was kept. Further, low

frequency levels (<1%) of binary variables were

investigated and 17 (constant or almost constant)

variables were removed. The redundancy of

categorical variables was checked using Cramer’s V

coefficient. Two variables with Cramer’s V

association with other variables higher than 0.8 were

removed. The final prepared data included 32

numerical and 15 categorical variables. The summary

statistics for demographic and some diagnosis, vital

parameters and patient reported outcomes in the

prepared data used for modelling are given in Tables

1 and 2 for RIBECCA and RIBANNA, respectively.

Table 1: Baseline characteristics of RIBECCA patients.

Variable

Count non-

missin

values

Mean (std)

Age (years) 500 63.8 (11.6)

Body-mass index

(

)

498 26.5 (4.9)

Days since primary

diagnosis

428

2234.7

(2373.2)

ECG QT interval

(ms)

497 384.5 (32.9)

ECG QRS interval

(

)

493 87.7 (17.9)

ECG PR interval

(

)

467 156.1 (25.6)

ECG heart rate

(

eats per minute)

497 74.56 (12.1)

EORTC physical

functioning revised

[0,100]

472 26.9 (23.5)

EORTC breast

symptoms [0,100]

464 13.9 (19.1)

Prediction of QT Prolongation in Advanced Breast Cancer Patients Using Survival Modelling Algorithms

167

Table 2: Baseline characteristics of RIBANNA patients.

Variable

Count non-

missing values

Mean(std)

Age (years) 1840 64.3 (11.6)

Body-mass index

(

)

1695 27.1 (5.7)

Days since primary

dia

nosis

1835

2171.2

(

2667.7

)

ECG QT interval

(ms)

1179 385.6 (33.9)

ECG QRS interval

(

)

1152 88.9 (15.9)

ECG PR interval

(

)

1045 156.1 (29.5)

ECG heart rate

(

eats per minute)

1233 77.5 (13.5)

EORTC physical

functioning revised

[0,100]

1407 35.5 (26.5)

EORTC breast

symptoms [0,100]

1358 17.3 (21.2)

4 METHODOLOGY

4.1 Survival Modelling Algorithms

In this study we applied and compared five survival

modelling algorithms: well-known statistical Cox

proportional hazards model (CPH), Cox proportional

hazards model regularized by elastic net (CPHNet),

gradient boosting survival model (GBS), random

survival forest (RSF) and fast survival support vector

machines (SSVM). A guide and references to these

algorithms can be found in the documentation of the

scikit-survival Python package (Pölsterl, 2020),

which we used in our study.

CPH is a type of regression model commonly used

in survival analysis to (1) estimate the risk of an event

over time and (2) identify predictive factors. It models

the hazard function assuming that input variables

(covariates) can affect the risk (i.e. hazard)

proportionally, i.e. the effect magnitude is time-

invariant. In other words, the initial difference in risk

of event for two patients remains constant over time.

Despite this restrictive assumption, CPH became a

very popular model due to its simplicity and

understandable output. Its major drawbacks however

are inability to perform in high-dimensional problems

with non-linear or interaction effects and correlated

features. Similarly to linear or logistic regression, the

latter issue can be mitigated by implementing and

optimizing the L2 shrinkage parameter in its loss

function.

CPHNet is an extension of CPH which

implements elastic net regularization that makes a

trade-off between L1 and L2 shrinkage. This

improves the numerical stability of the algorithm,

making it applicable to highly dimensional and

correlated problem settings. The issues with

modelling interactions and non-linearities remain the

same as in CPH. Survival machine learning

algorithms are developed to mitigate these issues.

GBS works similarly like the conventional

gradient boosting algorithm. It sequentially builds

multiple base learners (commonly regression trees),

which perform slightly better than random guessing.

These are called weak learners. Each weak learner

reduces the bias error by focusing on previously

inaccurately predicted learning examples (in our case

patients). In this way, the performance of the whole

additive model is boosted. The algorithm is trained in

a greedy manner, i.e. previously trained trees are

never revised and adjusted. Commonly optimized

hyperparameters are depth of base regression trees

and learning rate, which controls the contribution of

each tree to the overall prediction. The only

difference of GBS to its conventional counterpart is

introduction of the partial likelihood function of CPH

in its loss function, enabling it to model survival

functions.

RSF is a survival machine learning counterpart of

the conventional random forest algorithm, well-

known for its ability to reduce variance error. It trains

multiple decision trees on subsets of learning

examples and variables in parallel. The overall

prediction is obtained by aggregating trees’ outputs.

Analogue to GBS, the distinctive characteristic of

RSF comparing to conventional random forest is the

tree splitting criterion. Different splitting criteria have

been proposed to split tree nodes in branches with

different event times. One of the most popular criteria

is the log-rank test that was used in our study as well.

Hyperparameters of RSF that are typically tuned are

number of trees and max tree depth.

SSVM is an adaption of the conventional support

vector machine algorithm to model censored time to

event data. SSVM also employs a kernel function to

map input variable space into high-dimensional

feature space, where a hyperplane is fitted to

maximize the margin between examples (i.e. patients)

with dissimilar times to event. In our study we used

an efficient implementation of SSVM, testing

different kernel functions. Like in linear CPH and

CPHNet models, regularization strength

hyperparameter is typically optimized in SSVM as

well.

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

168

4.2 Performance Metrics

The standard performance metric for survival models

is the concordance index, also called Harrell’s c-

index or c-statistic. It quantifies how well the model

orders patients by their survival times (or times to

event), i.e. it estimates the probability that a patient

with higher predicted risk score is the one who

survives shorter, for each random pair of patients.

Analogue to the area under the receiver operating

characteristic curve in binary classification tasks, a c-

index of 0.5 indicates random guessing, while c-index

of 1 indicates perfect ordering of patients.

As shown in (Uno et al., 2011), c-index expresses

inflated, overly optimistic performance in problems

with increasing amount of censoring. The percentage

of censored examples is higher than 90% in both

RIBECCA and RIBANNA data, as stated in section

3.2.2. Therefore, we decided to use a version of c-

index based on Inverse Probability of Censoring

Weights (IPCW). IPCW assigns higher weights to

examples that are more likely to be observed, making

the estimate unbiased for this population. IPCW-

based c-index is then computed like a regular c-index,

taking IPCW weights into account.

4.3 Machine Learning Optimization

and Validation Pipeline

The machine learning pipeline included different

transformers for numerical and categorical variables.

Missing values in numerical variables were imputed

using iterative imputer based on Bayesian ridge

regression model (Bishop, 2006). Each variable with

missing values was modelled as a function of other

variables. For categorical variables, missing values

were imputed using simple imputer based on most

frequent value followed by dummy encoding (also

called one hot encoding), which created one binary

variable for each category. The machine learning

pipeline finally included a survival modelling

algorithm. Hyperparameters of included algorithms

were optimized in a grid search procedure using a 2-

fold cross-validation. The overview of optimized

hyperparameters is given for each algorithm in Table

3. To objectively assess model performance in the

internal validation (i.e. separately within RIBECCA

and RIBANNA) and avoid data leakage while

optimizing hyperparameters, another, outer 3-fold

cross-validation was implemented. This procedure

resulted in 3x2-fold nested cross validation (Cawley

et al., 2010). In the external, cross-study validation,

the outer cross-validation is excluded. All data

Table 3: Optimized hyperparameters for each algorithm.

orith

arameters

CPH

Regularization strength

alpha

CPHNet

Elastic net ratio between

L1 and L2 shrinka

GBS

Learning rate

Max. tree depth

RSF

Number of trees

Max. tree de

SSVM

Regularization strength

alpha

Kernel function (linear,

polynomial, radial basis

function

)

from one study was used for model training with

hyperparameter optimization and the model trained

with the best values of hyperparameters was applied

to another study.

4.4 Model Inspection

To enable model inspection and asses the importance

of included variables for the model performance, we

applied permutation feature importance method

(Breiman, 2001). This model-agnostic method

estimates how much the performance decreases when

a feature is randomly shuffled, i.e. not available in the

analysis. Feature importance is assessed only for the

best model in external, cross-study validation for both

studies.

5 RESULTS

5.1 Model Performance

As described in section 4.3, we performed internal,

nested cross-validation within each study as well as

external, cross-study validation. The performance

scores (IPCW-based c-index) of the former are shown

in Table 4 for each model. Moderate performance is

demonstrated by most models.

Table 4: Performance score (IPCW-based c-index) in

internal, nested cross-validation shown as mean (std).

Model RIBECCA RIBANNA

CPH 0.66

(

0.04

)

0.66

(

0.00

)

CPHNet 0.68 (0.05) 0.67 (0.02)

GBS 0.64 (0.09) 0.56 (0.07)

RSF 0.64 (0.07) 0.52 (0.10)

SSVM 0.51

(

0.02

)

0.65

(

0.02

)

Prediction of QT Prolongation in Advanced Breast Cancer Patients Using Survival Modelling Algorithms

169

Table 5: Performance score (IPCW-based c-index) in

external, cross-study validation.

Model

Training on

RIBECCA,

validation on

RIBANNA

Training on

RIBANNA,

validation on

RIBECCA

CPH 0.64 0.64

CPHNet 0.63 0.71

GBS 0.57 0.79

RSF 0.59 0.88

SSVM 0.58 0.60

Linear regularized CPHNet models reached the

highest score in both RIBECCA and RIBANNA

(0.68 and 0.67, respectively). The performance scores

in the external, cross-study validation are given in

Table 5. CPHNet showed relatively stable

performance. When trained on RIBECCA and tested

on RIBANNA, CPHNet reached the validation score

of 0.63. However, when trained on RIBANNA, it

reached notably higher validation score of 0.71 on

RIBECCA. GBS, RSF and SSVM also performed

better when trained on larger real-world RIBANNA

data and validated on smaller high quality, RIBECCA

trial data.

5.2 Predictive Factors

We were also interested in identifying the most

predictive factors of QT prolongation. For this

purpose, we applied the permutation feature

importance method described in section 4.4 to the

CPHNet model, which demonstrated the most

consistent performance across all validations. Figure

4 shows the top five features (all at baseline) of the

CPHNet model trained on RIBANNA and validated

on RIBECCA. The strongest predictor is the QT

interval in patient’s ECG at baseline. Other important

predictors include days since primary diagnosis, age,

and scores from two quality of life questionnaires.

Similarly, feature importance was also computed in

RIBANNA validation set, after training CPHNet on

RIBECCA (Figure 5). Baseline QT interval in ECG

again showed to be the most important predictive

factor, followed by heart rate, physical functioning

score, days since primary diagnosis and QRS interval

in patient’s ECG. Interestingly, vital signs (ECG

features) as well as patient quality of life (EORTC

features) dominate the top five features in both

evaluations. It should be noted that permutation

feature importance was based on models with limited

performance (especially when trained on RIBECCA

and validated on RIBANNA) and therefore should be

interpreted with care.

Figure 4: Feature importance for model trained on

RIBANNA and validated on RIBECCA.

Figure 5: Feature importance for model trained on

RIBECCA and validated on RIBANNA.

6 CONCLUSION AND FUTURE

WORK

In this paper we presented the feasibility of predicting

QT prolongation in HR+/HER2- advanced breast

cancer patients treated with CDK4/6 inhibitor

ribociclib using survival modelling algorithms. We

trained and compared the performance of five

statistical and machine learning algorithms for

survival analysis, observing that Cox proportional

hazards model regularized by elastic net (CPHNet)

demonstrated the most consistent performance,

mostly higher than the performance of the well-

known statistical Cox proportional hazards model

(CPH). Models trained on the clinical trial data

(RIBECCA) showed moderate performance when

validated on the real-world data (RIBANNA). This is

most likely due to lower real-world data quality

(many more missing values which needed to be

imputed during testing) and higher data variety,

which is not properly captured by models trained on

small trial data only. In addition, since ranges of

numerical variables in RIBANNA are larger than in

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

170

RIBECCA, models were sometimes extrapolating

when validated on RIBECCA, contributing to the

performance loss. On the other hand, once trained on

larger, real-world RIBANNA data, models were

performing relatively well on high quality trial data

(IPCW-based c-index of the best model was 0.88, see

Table 5).

In addition to performance comparison, the most

predictive factors were identified in both studies,

when used for external validation. Whilst based on

imperfect models and thus interpreted cautiously, the

strongest predictors mostly include baseline ECG

variables (like QT interval) and EORTC patient

quality of life scores, in addition to days since

primary diagnosis and age. None of the cancer

severity features, prior therapies or hormone status

appeared among the top five predictive factors for QT

prolongation.

Based on these results, we strongly believe that

the presented methodology would be useful in a wide

range of tasks aiming at prediction of clinical events

and their times. In the future, we plan to tackle

modelling of further tumour control and safety

outcomes like progression-free survival or different

toxicities in cancer patients. Furthermore, we aim to

incorporate explainable AI approaches like SHAP

(Lundberg et al., 2017) to enable deeper insights into

predictive factors and explain predictions for

individual patients.

REFERENCES

Arnold M., Morgan, E., Rumgay, H., et al. (2022). Current

and future burden of breast cancer: Global statistics for

2020 and 2040. In The Breast, volume 66, pages 15-23.

Lu Y.S., Im S.A., Colleoni, M., et al. (2022). Updated

overall survival of ribociclib plus endocrine therapy

versus endocrine therapy alone in pre- and

perimenopausal patients with HR+/HER2- advanced

breast cancer in MONALEESA-7: A phase III

randomized clinical trial. In Clin Cancer Res, volume

28, issue 5, pages 851-859.

Ward, M., Harnett, J., Bell, T.J., et all. (2019). Risk factors

of QTc prolongation in women with hormone receptor‒

positive/human epidermal growth factor receptor 2‒

negative metastatic breast cancer: A retrospective

analysis of health care claims data. In Clinical

Therapeutics, volume 41, number 3, pages 494-504.

Brody, T. (2016). Clinical Trials: Study Design, Endpoints

and Biomarkers, Drug Safety, and FDA and ICH

Guidelines, Academic Press. London, 2

edition.

Spooner, A., Chen, E., Sowmya, A., et al. (2020). A

comparison of machine learning methods for survival

analysis of high-dimensional clinical data for dementia

prediction. In Sci Rep, 10, 20410.

Qiu, X., Gao, J., Yang, J., et al. (2020). A comparison study

of machine learning (random survival forest) and

classic statistic (cox proportional hazards) for

predicting progression in high-grade glioma after

proton and carbon ion radiotherapy. In Front. Oncol.

10:551420.

Moncada-Torres, A., van Maaren, M.C., Hendriks, M.P. et

al. (2021). Explainable machine learning can

outperform Cox regression predictions and provide

insights in breast cancer survival. In Sci Rep, 11, 6968.

Lyon, A.R., Dent, S., Stanway, S., et al. (2020). Baseline

cardiovascular risk assessment in cancer patients

scheduled to receive cardiotoxic cancer therapies: a

position statement and new risk assessment tools from

the Cardio-Oncology Study Group of the Heart Failure

Association of the European Society of Cardiology in

collaboration with the International Cardio-Oncology

Society. In European Journal of Heart Failure, 22,

pages 1945-1960.

Tisdale, J.E., Jaynes, H.A., Kingery, J.R., et al. (2013).

Development and validation of a risk score to predict

QT interval prolongation in hospitalized patients. In

Circ Cardiovasc Qual Outcomes, volume 6, number 4,

pages 479-487.

Haugaa, K.H., Bos, J.M., Tarrell, R.F., et al. (2013).

Institution-wide QT alert system identifies patients with

a high risk of mortality. In Mayo Clinic Proceedings,

volume 88, issue 4, pages 315-325.

Vandael, E., Vandenberk, B., Vandenberghe, J., et al.

(2017). Development of a risk score for QTc-

prolongation: The RISQ-PATH study. In Int. J. Clin.

Pharm, volume 39, pages 424–432.

Vandael, E., Vandenberk, B., Vandenberghe, J., et al.

(2018). A smart algorithm for the prevention and risk

management of QTc prolongation based on the

optimized RISQ-PATH model. In Br J Clin Pharmacol,

volume 84, pages 2824–2835.

Fasching, P.A., Wöckel, A., Tesch, H., et al. Machine

learning to predict treatment response and tolerability

in HR+, HER2– advanced breast cancer: German study

AI4ANNA. In San Antonio Breast Cancer Symposium

(SABCS 2022), abstract.

Decker T., Fasching, P.A., Nusch, A., et al. (2021). Efficacy

and safety of ribociclib (RIB) in combination with

letrozole (LET) in patients with estrogen receptor–

positive advanced breast cancer (ABC): Secondary and

exploratory results of phase 3b RIBECCA study. In

Annals of Oncology (ESMO), abstract 247P (poster).

Lüftner D., Brucker, C., Decker, T., et al. (2022). Real-

world efficacy of ribociclib (RIB) plus aromatase

inhibitor (AI)/fulvestrant (FUL), or endocrine

monotherapy (ET), or chemotherapy (CT) as first-line

(1L) treatment (tx) in patients (pts) with hormone

receptor–positive (HR+), human epidermal growth

factor receptor-2–negative (HER2–) advanced breast

cancer (ABC): Results of fourth interim analysis (IA)

from RIBANNA. In J Clin Oncol, abstract 1065

(poster).

Prediction of QT Prolongation in Advanced Breast Cancer Patients Using Survival Modelling Algorithms

171

Roden, D.M. (2016). Predicting drug-induced QT

prolongation and torsades de pointes. In J Physiol

volume 594, issue 9, pages 2459-2468.

Schwartz, P.J., Woosley, R.L. (2016). Predicting the

unpredictable: drug-induced QT prolongation and

torsades de pointes. In JACC, volume 67, number 16,

pages 1639-1650.

Pölsterl, S. (2020). “scikit-survival: A Library for time-to-

event analysis built on top of scikit-learn,” In Journal

of Machine Learning Research, volume 21, number

212, pages 1–6.

Uno, H., Cai, T., Pencina, M.J., et al. (2011). On the C-

statistics for evaluating overall adequacy of risk

prediction procedures with censored survival data. In

Statist. Med., volume 30, issue 10, pages 1105-1117.

Bishop, C. (2006). Pattern Recognition and Machine

Learning, Springer-Verlag. New York, 1

edition.

Cawley, G.C., Talbot, N.L.C. (2010). On over-fitting in

model selection and subsequent selection bias in

performance evaluation. In J Mach Learn Res., volume

11, pages 2079–2107.

Breiman, L. (2001). Random forests, In Machine Learning,

volume 45, number 1, pages 5-32.

Lundberg, S.M., Lee, S.I. (2017). A unified approach to

interpreting model predictions. In Adv Neural Inf

Process Syst, volume 30, pages 4768–4777.

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

172