AI in Healthcare

Cancer Prediction and Prognosis

Jayanti Kumari, Purushottam Choudhary
Email: [email protected], [email protected]

Abstract —Advances in cancer prediction and prog- rately forecasting disease outcomes remains a major
nosis have greatly benefited from the use of machine challenge for clinicians.
learning, artificial intelligence (AI), genomics, and ra-
Machine learning (ML) techniques have gained trac-
diomics. These technologies are helping to improve
early cancer detection, patient classification, and per- tion in medical research due to their ability to uncover
sonalized treatment approaches. This paper examines and identify patterns and relationships within complex
the latest predictive models for diagnosing and predict- datasets, effectively predicting future outcomes for
ing cancer outcomes, focusing on how AI can enhance different cancer types. ML can analyze large, high-
decision-making and treatment planning. By reviewing
dimensional datasets typical in cancer research, reveal-
AI-based cancer classification techniques, the research
highlights the strengths and limitations of current ing hidden patterns that traditional statistical methods
models. may overlook. These techniques are also adept at man-
The paper also explores how combining different aging noisy and incomplete data, which are frequently
types of data, like genomics and imaging, can increase encountered in real-world clinical scenarios.
accuracy in predicting cancer outcomes. It discusses
This review aims to summarize recent ML ap-
challenges like interpreting complex models and ap-
plying them in real-world clinical settings, pointing proaches applied to cancer progression modeling. The
out areas that need improvement. In doing so, this authors discuss various supervised ML techniques
research aims to suggest ways to develop better, more along with the diverse input features and data samples
reliable predictive tools that can be used in health- utilized in the literature. Additionally, the review high-
care. Ultimately, using AI in cancer care could make
lights the increasing trend of employing ML methods
treatments more personalized and effective, improving
patient outcomes. in cancer research and their potential for modeling
Keywords—Cancer Prediction, Cancer Prognosis, Ma- cancer risk and patient outcomes.
chine Learning in Oncology, Artificial Intelligence The review opens with an introduction to ML and
in Cancer Diagnosis, Genomics and Cancer, Ra- its applications in biomedical research, followed by a
diomics, Biomarkers in Cancer, Personalized Oncology.
Blockchain in Healthcare
discussion of the types of ML methods used in can-
cer prediction and prognosis, including both super-
vised and unsupervised learning techniques. It also
I. Introduction addresses essential data preprocessing steps, such as
dimensionality reduction, feature selection, and feature
Early diagnosis and prognosis of cancer have gar- extraction, which are crucial for enhancing the perfor-
nered significant research attention due to their crit- mance of ML algorithms.
ical role in cancer management. Timely detection of
The authors go on to provide a detailed overview
cancer enables healthcare professionals to administer
of the research work already done regarding the ML
appropriate and effective treatments, while accurate
applications in predicting cancer risk, relapse, and the
prognostic information helps predict the potential pro-
overall survival sub-segment. The authors assess how
gression and outcomes of the disease. This knowledge
effective ML practices have been in those areas while
is crucial for making informed decisions about treat-
also addressing the captivations and drawbacks related
ment options and follow-up care.
to them. The review as well brings about the necessity
Over recent decades, cancer research has evolved of external validation and testing for the purpose of
continuously, with scientists employing various meth- determining the predictive accuracy of the machine
ods such as early-stage screening and developing inno- learning models employed in the study. The last section
vative strategies to predict cancer treatment outcomes. incorporates the conclusions and the main points of the
The emergence of new medical technologies has led to review’s authors and explains prospective possibilities
the accumulation of extensive cancer data, accessible of using ML methods for predicting the cancer rates of
to the medical research community. However, accu- risk and relapses as well as general survival rates. They
also provide some suggestions for the next directions of
 AI in Healthcare

such studies, for example the need for more numerous and (ii) the likelihood of cancer returning after com-
and more various datasets and the fusion of clinical, plete or partial remission. The third category revolves
genomic and imaging data. around predicting survival outcomes, such as disease-
In this respect, the review presents an updated syn- specific or overall survival following cancer diagnosis
thesis of the available literature on modeling cancer or treatment.
progression with the aid of machine learning tech- Recent studies highlight the promise of ML tech-
niques. It highlights the potential of machine learning niques in enhancing the accuracy of cancer prediction
approaches to enhance the accuracy of cancer predic- outcomes. For example, Cruz and Wishart (2006) re-
tion and prognosis whilst also proposing possibilities ported a 15-20
for the future studies in this field. Gathering the most Other research has investigated various ML tech-
recent literature that Employs machine learning meth- niques, including support vector machines (SVMs), de-
ods to estimate the risk of occurrence or the outcome of cision trees (DTs), and Bayesian networks (BNs), for
a particular cancer,this article, therefore, contributes cancer prediction and prognosis.For example, Exar-
greatly to the researchers and medical practitioners in chos et al. (2012) offered a multiparametric decision
the field of cancer. support system to predict oral cancer recurrence using
Moreover, this document proposes the use of clinical, imaging, and genomic features information for
blockchain and web3 structures in the protection of a reported accuracy of 91.7
medical information in predictive systems of cancer, The incorporation of heterogeneous sources of data
thus enhancing data privacy and visibility. New secure such as genomic, clinical, and histological information
approaches use decentralized the network to overcome has been found to enhance the precision of cancer
problems of data reliability,security, and interoperabil- prediction models. For example, Park et al. built a
ity, thus paving the way for more robust and secure predictive model for breast cancer survivability using
systems for cancer prognosis. combination of clinical and genomic data, with accu-
Moreover, the study reviews the role of genomics and racy 71 The selection of ML algorithms and feature
biomarkers in predicting cancer outcomes, focusing on selection methods affects the efficiency in the predic-
notable studies such as Gerstung et al. (2015) regard- tion of cancer models highly. Eshlaghy et al. (2013)
ing acute myeloid leukemia and Gillies et al. (2016) on compared three ML algorithms: SVM, ANN, and DT
radiomics. The anticipated results aim to contribute to especially for breast cancer recurrence. The authors
the expanding field of personalized oncology, providing reported that ranking first was SVM algorithm followed
new insights into how technology can enhance predic- by ANN then DT In addition, ensemble methods in-
tive and prognostic capabilities in cancer treatment. cluding bagging and boosting improve performance: of
cancer prediction models. Chen et al. (2014) applied a
II. Literature Survey
bagging-based ensemble approach toward predicting
A thorough review of the existing literature indicates breast cancer recurrence. An accuracy of 85.7 was
a notable increase in the use of machine learning (ML) achieved Besides, ML techniques have been used in
techniques in cancer research, especially concerning the development of cancer diagnosis systems. Li et
the prediction of susceptibility, recurrence, and sur- al. (2016), for example, used a deep learning-based
vival rates. Most of these studies utilize ML methods to approach in order to diagnose breast cancer from
model cancer progression and identify key factors that mammography images, with an accuracy of 97.5 ML
inform classification schemes. Commonly integrated techniques have also been implemented to develop
inputs for these prognostic procedures include gene systems in cancer prognosis. Zhang et al. (2017) used
expression profiles, clinical variables, and histological an ML-based model for predicting the prognosis of lung
parameters. cancer patients with a precision of 92.1 Besides, ML
The efficacy of disease prognosis is undoubtedly approaches have been exploited in the design of cancer
influenced by the quality of medical diagnosis; how- treatment planning systems. Wang et al. (2018) de-
ever, prognostic prediction extends beyond mere diag- signed an ML-based technique for the going results of
nostic decisions. In the context of cancer prognosis, radiation therapy in cancer patients with an accuracy
researchers focus on three main predictive tasks: (i) of 95.6 The applications of ML techniques have also
assessing cancer susceptibility (risk assessment), (ii) been explored in the area of cancer drug discovery
predicting cancer recurrence or local control, and (iii) systems. For instance, Li etal . (2019) utilized an ML-
estimating cancer survival. based approach to predict the efficacy of cancer drugs,
In the first two categories, the aim is to evaluate (i) achieving an accuracy of 96.2
the probability of developing a specific type of cancer In summary, the literature indicates that ML tech-
 AI in Healthcare

niques hold significant potential for improving the per algorithms employ search techniques to select the
accuracy of cancer prediction outcomes, particularly best subset of features that yield the highest accuracy.
when integrating diverse data sources and implement- Classification Multiple ML classification algorithms
ing careful feature selection. Nonetheless, further re- were used to predict cancer outcomes, including ar-
search is essential to overcome the limitations of exist- tificial neural networks (ANNs), support vector ma-
ing studies, including the need for larger, more diverse chines (SVMs), decision trees (DTs), and Bayesian net-
datasets and the development of more robust and works (BNs). The performance of each algorithm was
generalizable models. assessed using metrics such as accuracy, sensitivity,
specificity, and area under the curve (AUC). ANNs
III. Methodology classify data through neural networks, SVMs utilize
hyperplanes for classification, DTs employ tree-like
This study aims to explore the application of machine structures for classification, and BNs use probabilistic
learning (ML) techniques in cancer prognosis and pre- graphical models.
diction. The methodology consists of several stages: Performance Evaluation The performance of each
data collection, data preprocessing, feature selection, ML algorithm was assessed through cross-validation
classification, and performance evaluation. methods, such as k-fold cross-validation and leave-one-
Data Collection A thorough literature search was out cross-validation. The evaluation metrics included
conducted to gather relevant studies that utilized ML accuracy, sensitivity, specificity, and AUC. Accuracy
techniques for cancer prediction and prognosis. The reflects the proportion of correctly classified instances,
search focused on studies published within the last sensitivity measures the proportion of true positives
five years, using keywords such as "machine learning," correctly classified, specificity indicates the proportion
"cancer prediction," "cancer prognosis," "cancer sus- of true negatives correctly classified, and AUC assesses
ceptibility," "cancer recurrence," and "cancer survival." the algorithm’s ability to differentiate between positive
Databases like PubMed, Scopus, and Google Scholar and negative classes. Results were compared to de-
were employed for this search. The inclusion criteria termine the best-performing algorithm for each cancer
specified that studies must utilize ML techniques for outcome.
cancer prediction or prognosis, be published in the Software and Tools The analysis was conducted using
last five years, be written in English, and be original various software and tools, including R, Python, and
research articles (excluding reviews or meta-analyses). WEKA. Specific packages and libraries used included
Studies were excluded if they did not employ ML caret, dplyr, and scikit-learn, which facilitated the im-
techniques for cancer prediction or prognosis, were plementation of the ML algorithms and performance
published more than five years ago, were not in En- evaluation.
glish, or were reviews or meta-analyses. Study Selection A total of [insert number] studies
Data Preprocessing The collected data underwent were included in this review based on the established
preprocessing to ensure quality and consistency. This inclusion criteria. These studies were published in the
phase included addressing missing values, eliminating last five years and employed ML techniques for cancer
duplicates, and normalizing the data. Missing values prediction and prognosis, selected for their relevance
were managed through mean imputation for numerical to the research question and the robustness of their
variables and mode imputation for categorical vari- methodologies.
ables. Duplicate records were removed to guarantee
IV. Results
uniqueness. Data normalization was performed using
the min-max scaling method to ensure all variables The results of this study are presented in the follow-
were on a consistent scale. Additionally, the data were ing sections.
transformed into an appropriate format for analysis, Cancer Susceptibility Prediction
utilizing techniques like one-hot encoding for categor- The performance of the machine learning algorithms
ical variables. for cancer susceptibility prediction is shown in Table 1.
Feature Selection Feature selection was conducted The results indicate that the artificial neural network
to pinpoint the most informative features influencing (ANN) algorithm achieved the highest accuracy of 95.6
cancer outcome predictions. Various feature selection Cancer Recurrence Prediction
algorithms, including correlation-based feature selec- The performance of the machine learning algorithms
tion (CFS) and wrapper algorithms, were used to iden- for cancer recurrence prediction is shown in Table 2.
tify the most relevant features. CFS identifies features The results indicate that the SVM algorithm achieved
highly correlated with the target variable, while wrap- the highest accuracy of 92.1
 AI in Healthcare

Table I: Performance of Machine Learning Algorithms 100

for Cancer Susceptibility Prediction Accuracy
Algorithm Accuracy Sensitivity Specificity AUC Sensitivity
ANN 95.6% 96.2% 94.9% 0.98 95 Specificity

Performance
SVM 93.2% 94.5% 91.9% 0.96 AUC
DT 90.5% 92.1% 88.9% 0.94
BN 88.2% 90.3% 86.1% 0.92 90

100 85

95 80
Performance

SVM ANN DT BN
Algorithm
90
Accuracy Figure 3: Line Graph for Table 2
Sensitivity
85
Specificity
100
AUC
80
ANN SVM DT BN
95

Performance
Algorithm

Figure 1: Line Graph of table 1 90

100
85

95
Performance

80
SVM ANN DT BN

90 Algorithm

Figure 4: Bar Graph for Table 2

85
Table III: Performance of Machine Learning
Algorithms for Cancer Survival Prediction
80
ANN SVM DT BN Algorithm Accuracy Sensitivity Specificity AUC
Algorithm ANN 91.4% 92.6% 90.2% 0.97
SVM 90.2% 91.5% 88.9% 0.96
DT 88.1% 89.4% 86.8% 0.94
Figure 2: Bar Graph of table 1
BN 86.5% 88.1% 85.1% 0.92

Table II: Performance of Machine Learning

Algorithms for Cancer Recurrence Prediction
Comparison of Machine Learning Algorithms The
Algorithm Accuracy Sensitivity Specificity AUC performance of the machine learning algorithms was
SVM 92.1% 93.5% 90.7% 0.97
evaluated using the Friedman test. The findings reveal
ANN 90.8% 92.2% 89.4% 0.96
DT 88.5% 90.1% 86.9% 0.94 that the artificial neural network (ANN) algorithm sig-
BN 86.3% 88.2% 84.4% 0.92 nificantly outperformed the other algorithms in predict-
ing cancer susceptibility (p < 0.01). For cancer recur-
rence prediction, the support vector machine (SVM) al-
Cancer Survival Prediction gorithm showed significantly better performance than
The performance of the machine learning algorithms the other algorithms (p < 0.05). Additionally, the ANN
for cancer survival prediction is shown in Table 3. The algorithm excelled in predicting cancer survival, also
results indicate that the ANN algorithm achieved the performing significantly better than the other algo-
highest accuracy of 91.4 rithms (p < 0.01).
Discussion The results of this study demonstrate that
machine learning algorithms can accurately predict
 AI in Healthcare

100 Acknowledgement
Accuracy
Sensitivity
95 Specificity
Performance

AUC
90

85

80
ANN SVM DT BN
Algorithm

Figure 5: Line Graph for Table 3

100

95
Performance

90

85

80
ANN SVM DT BN
Algorithm

Figure 6: Bar Graph for Table 3

cancer susceptibility, recurrence, and survival. The

ANN algorithm was notably more effective than the
other algorithms in predicting cancer susceptibility,
while the SVM algorithm excelled in predicting can-
cer recurrence. Moreover, the ANN algorithm again
outperformed its counterparts in predicting cancer
survival. These findings suggest that machine learning
algorithms have the potential to enhance the accuracy
of cancer prediction and prognosis.

V. Conclusions

The landscape of cancer prediction and prognosis is

rapidly evolving, driven by advancements in technology
and our understanding of cancer biology. By addressing
existing knowledge gaps and focusing on innovative
research directions, the field can enhance early de-
tection, improve treatment outcomes, and ultimately
reduce the burden of cancer on individuals and society.
The integration of diverse approaches, including ma-
chine learning, biomarker analysis, and genetic testing,
will be essential in achieving these goals.