Application of machine learning methods to predict progression in patients with hormone-sensitive prostate cancer

doi:10.21203/rs.3.rs-3309939/v1

Objective

Precise and appropriate diagnosis for prostate cancer patients can improve their quality of life. We sought to develop an innovative machine learning prognostic model to forecast the progression of hormone-sensitive prostate cancer (mHSPC).

Methods

A retrospective cohort study was conducted at Yunnan Cancer Hospital, including 533 patients diagnosed with hormone-sensitive prostate cancer between January 2017 and February 2023.In this machine learning model, K-proximity algorithm (KNN), naive Bayes, random forest algorithm, XGBoost and ADAboost were used to establish prediction models. The main evaluation indicators were the accuracy(ACC), precision༈PRE༉, specificity༈SPE༉, sensitivity༈SEN༉or regression rate ༈Recall༉and f1 score of the model.

Results

We established KNN, Naive Bayes, random forest algorithm, XGBoost and ADAboost models, and their accuracy rates were 75.4%, 71.1%, 88.02%, 86.6% and 85.2%, respectively.Among the generated models, XGboost has the highest accuracy of 88.02%.

Conclusion

Our model is more accurate and perfect than the predecessors, and can provide reference for clinical work.

Biological sciences/Cancer

Health sciences/Medical research

Machine learning

hormone sensitivity

prostate cancer prediction model

Prostate cancer is the second most prevalent cancer globally, following lung cancer. The incidence and mortality rates of prostate cancer are strongly correlated with advancing age^[1].Prostate-specific antigen (PSA) is extensively utilized as the primary biomarker for prostate cancer screening^{[2, 3]}.Studies have shown that gleason score, PSA, serum markers and other indicators play a crucial role in the prognosis of patients with prostate cancer^[4].Clinical treatment decisions are often made around these indicators. However, prostate cancer experts do not determine the severity of each indicator of prostate cancer progression^[5],which creates difficulties for clinical work。

Artificial Intelligence (AI) is considered a field that focuses on automating intelligent tasks normally performed by humans. And Machine Learning (ML) is the way to do it^[6].Machine learning has already become a pivotal component in various applications, including text detection and recognition, early prediction, power quality disturbance detection etc^[7].In clinical practice, machine learning is mainly used for predictive analysis and optimization of clinical management decisions. At present, this technology has achieved good results in the detection of diseases such as diabetic retinopathy and breast disease, which has great potential in clinical work^[8].We believe that machine learning can also predict the progression of prostate cancer and guide our clinical work through its predicted results.

In this study, Python 3.8.6 was employed to construct prediction models for the progression of hormone-sensitive prostate cancer patients, and the significance of each factor in the final outcome was analyzed,aiming to build a relevant prediction model.Data was collected from patients diagnosed with hormone-sensitive prostate cancer at the Department of Urology in Yunnan Cancer Hospital, and constructed K-proximity algorithm (KNN), Random forest, Naive Bayes model, XGboost, and ADAboost models.Our goal is to find the best predictive model among them.

Patient inclusion criteria

Inclusion criteria：

Patients with hormone-sensitive prostate cancer (mHSPC) confirmed by histopathology
Imaging examination (B-ultrasound, CT, MRI, bone scan, etc.) showed measurable lesions
Those who can provide complete medical records and follow-up data

All patients received treatment with a luteinizing hormone-releasing hormone agonist (LHRHa) during their course of treatment.

Exclusion criteria

Prostate cancer patients without definite histopathological diagnosis
Patients with other malignant tumors
Patients with systemic diseases such as blood diseases and immune system, which affect hematological indicators
Patients who refuse or cannot cooperate with clinical data collection or follow-up due to other reasons

（5）Patients with other serious organ diseases

Patient characteristics

It included the patient's diagnosed age, TNM stage, Gleason score, testosterone value at first diagnosis, prostate volume size at first diagnosis, fPSA, TPSA and their ratio at first diagnosis, Tumor burden, whether there was bone metastasis and visceral metastasis, alkaline phosphatase at first diagnosis, treatment plan and other clinical data, as well as B-ultrasound, CT, PET-CT, MRI, and whole body bone scan Imaging and other imaging data.

A high tumor load was defined as having internal metastases or bone metastases >=4 sites, with at least one metastatic site located beyond the spinal column or pelvic region^[9].The progression of prostate cancer was defined as three consecutive serum PSA tests at an interval >=1 week, with a continuous increase in PSA and two consecutive increases of more than 50% from the baseline value (the lowest value). If nPSA<2ng/ml, the absolute value of the PSA increase should be greater than 2ng/ml.

Imaging progress was defined as the discovery of new lesions on imaging, including the presence of two or more new bone metastases detected on a bone scan or the appearance of new soft tissue lesions.

Patient outcome judgment

If the patient did not develop PSA progression and imaging progression, the last follow-up time was used as the cut-off point. The prerequisite for judging progress is the castration level of PSA, i.e. the castration level of serum testosterone after androgen deprivation therapy (ADT) (<1.7nmol/L or <50ng/dl).

Model development

Model construction

We input these collected patient profile sets into jupyter software, the tool for implementing python, which is an open-source, object-oriented, interpreted, and interactive programming language.The primary predictive outcome was whether the patient progressed over the course of a year.With scikit-learn, we divide the data, establishing a training set and a test set with an 8:2 distribution. In order to make the model establishment stable, we set the random number seeds used in the learning process to 42. The dataset had a disparity in the number of positive and negative results. As a solution, we decided to oversample all the data^[10].

KNN, Naive Bayes, random forest in ensemble learning, XGboost and ADAboost were adopted for model selection^[11-15].We also standardized the data before using KNN^[16].Parameter adjustment mainly relies on five-fold cross-validation and grid search to determine hyperparameters. In order to prevent overfitting of the model,we added Early Stopping strategies to improve our function.We also visualized the parameter adjustment process of ensemble learning, as shown in Figure(S3) and (S5) respectively.

All models were evaluated using accuracy (ACC), precision（PRE）, specificity（SPE）, sensitivity（SEN）or regression rate （Recall）and f1 score, where TP represented true positive, TN represented true negative, FP represented false positive, and FN represented false negative.Their specific performance is presented in the form of a confusion matrix. Python 3.8.6 is used for the entire machine learning.

According to the inclusion and exclusion criteria, we collected 826 newly diagnosed hormone-sensitive prostate cancer patients from Yunnan Cancer Hospital from January 2017 to January 2022, among which 293 missing data were excluded, and their problem data mainly focused on Gleason score, pathological type, and absence of PSA at initial diagnosis. Our final cohort consisted by 533 patients.

We modeled the data characteristics of these patients and adjusted the parameters described above. The characteristic subsets and baseline data of different patients can be seen in Table 1 and Table (S1).Figure 2 is a subset attribute heat map. The distribution of patient indicators is shown in Figure (S1).We used python3.8.6 to carry out Shapiro-Wilktest test method to test the normality of measurement data. For measurement data conforming to normal distribution, two independent sample T-test was used, and "mean ± standard deviation" was used for statistical description.Wilcoxon rank sum test was used to describe the data that did not conform to normal distribution, and "median (25%-75%) [M (P25-P75)]" was used for statistical description. Counting data and grade data are expressed by frequency (%); Chi-square test was used to compare the differences between groups of other bivariate variables.

Test level: α=0.05, P<0.05 was considered statistically significant.

Figure 1 shows our technical path.

Model representation

KNN model

KNN model accuracy (ACC) was 75.4%, sensitivity(SEN) or regression rate(Recall) was 80%, specificity (SPE) was 70.1%, precision（PRE） was 75%, and f1-score was 77.42%, as shown in Figure 3. The area under the ROC curve (AUC) is 0.75, as shown in Figure 4.

Naive bayes

The accuracy rate (ACC) of the naive Bayes model was 71.1%, the sensitivity (SEN) or regression rate (Recall) was 83.1%, the specificity (SPE) was 62.65%, the precision（PRE）was 61.25%, and the f1-score was 70.5%, as shown in Figure 5. The area under the ROC curve (AUC) is 0.76, as shown in Figure6.

Ensemble learning

The ensemble learning in order of model accuracy from large to small is: XGboost, ADAboost, random forest.The accuracy (ACC) of XGboost model was 88.02%, sensitivity(SEN) or the regression rate(Recall) was 90.9%, the specificity (SPE) was 84.6%, the precision (PRE) was 87.5%, and the f1-score was 89%.ADAboost model accuracy (ACC) was 86.6%,sensitivity(SEN) or regression rate (Recall) was 89.6%, specificity (SPE) was 83.1%, the precision（PRE） was 86.25%, and f1-score was 87.9%.The random forest model accuracy (ACC)was 85.2%,sensitivity(SEN) or regression rate(Recall) was 87.3%, specificity (SPE) was 82.5%, the precision（PRE） was 86.25%, and f1-score was 86.79%, as shown in Figure 7.The order of area under ROC curve (AUC) from highest to lowest is: ADAboost0.93, Random forest 0.92, XGboost0.89, as shown in Figure 8.

Analysis of importance based on ensemble learning

The importance selection of various features in the random forest is shown in Figure 9. Among them,tumor burden, hormone sensitive stage treatment plan, lactate dehydrogenase (LDH), alkaline phosphatase, and whether bone metastasis occurs at first diagnosis are in the top 5.At the bottom of the list were the presence or absence of hematuria, Gleason score and pathological type. The random forest model visualization is shown in Figure (S4).

The importance of features in the XGboost model is shown in Figure 10. The ranking of XGboost is calculated by the sum of error reduction in variable segmentation.LDH and prostate volume ranked first in the selection of importance of each feature, while bone metastasis and Gleason score ranked last. The XGboost model visualization is shown in Figure (S6).

Measures such as PSA, gleason score, and sex hormones have been shown to have a role in mHSPC patients, but for a long time there has been no conclusive evaluation of the importance of these measures^[5].Machine learning has aroused clinical interest because of its ability to provide predictive analysis of disease^[17].However, machine learning for prostate cancer is rare. A study on active monitoring of the progress of adenocarcinoma patients^[18]established a model similar to ours and evaluated it with f1-score, but the f1-score of each model was lower than 0.6.Another support vector machine (SVM) model has an accuracy of 0.852^[19].Our study complements this gap and establishes several models with higher accuracy than previous ones.

Data set characteristics have a considerable impact on the end result of machine learning.There is a study^[20]suggests that the model performance improves with increasing granularity and has advantages when new variables are included.Our study proves this by adding features such as prostate volume and hematuria and perfecting a more complete model construction compared to previous studies.However, the random forest model and XGboost model have different judgments on clinical factors, which may be due to their different working modes, which lead to their different selection of feature priorities ^{[13, 14]}.

In spite of this, the two still have some similarities in the feature rankings, such as lactate dehydrogenase (LDH) and alkaline phosphatase are relatively high in their feature rankings. Studies have shown their value in prostate cancer^{[21, 22]}.However, in clinical work and guideline recommendations, such as PSA and testosterone are still the first choice for follow-up, and there is even no evaluation of LDH level. We believe that the results of machine learning models can provide a reference for clinical work, but the validity of such results needs to be confirmed by subsequent clinical observation or clinical trials.

This study has several shortcomings. First, the data we collected were retrospective, and there were deficiencies such as data missing, bias and confounding factors. Secondly, we cannot rule out the influence of treatment scheme on the whole model. Due to the heterogeneity of variables, different patients have different responses to the same treatment scheme, and treatment scheme is a key variable in the random forest. In the future, we need to collect more perfect data and explore more influencing factors to build a better model which could guide clinical work.

Despite these limitations, we have improved the construction and accuracy of machine learning models for mHSPC patients based on previous work, and this study can provide a basis for stable model construction with larger sample sizes in the future.

On the basis of summarizing previous experience, this study improved the mHSPC patient progression model and created a more accurate model. We believe that this study can provide reference value in future clinical work.

mHSPC (hormone-sensitive prostate cancer),ACC（accuracy）, PRE（precision）, SPE（specificity）, SEN（sensitivity）, Recall（regression rate）,PSA (Prostate-specific antigen),AI (Artificial Intelligence)ML (Machine Learning),KNN (K-proximity algorithm),LHRHa (luteinizing hormone-releasing hormone agonist), ADT(androgen deprivation therapy),SVM (support vector machine)

Author Contributions

Yu Bai conceived and designed the study;Longguo Dai,Huijian Wang,Bingyu Zhu,Kun Zhang,Yang Wang,Feiyu Yin,Ji Li collected data;Qilin Wang,Hong Yang,Ruiqian Li,Yu Bai,Jun Li,Hongyi Wu,Chen Hu,Haiyang Jiang and Chongjian Zhang performed the clinical treatment of the prostate cancer;Bingyu Zhu and Haiyang Jiang wrote the main manuscript text and prepared the figures;All authors have read andapproved the manuscript.

Funding

The retrospective study was supported by National Natural Science Foundation of China（No.82160511）,

National Cancer Center Climbing Fund（No. NCC201925B01）,

Young and Middle-aged Academic and Technical Leaders Reserve Talent Program of Yunnan Province (202305AC160053)

Ethics approval and consent to participate

The study protocol was approved by the Yunnan Cancer Hospital Ethics Committee. All methods were carried out in accordance with relevant guidelines

and regulations. All recipients signed informed consent and for dead patients

informed consent from legal guardians/next of kin were obtained.

Availability of data and materials

The datasets used and analysed during the current study available from the corresponding author on reasonable request.

Informed Consent Statement

Informed consent was obtained from all patients involved in the study.

Data Availability Statement

The original contributions presented in the study are included in the article/Supplementary Material. Further inquiries can be directed to the corresponding authors.

Conflicts of Interest

The authors declare no conflict of interest.

Rawla, P., Epidemiology of Prostate Cancer. World J Oncol, 2019. 10(2): p. 63–89.
Duffy, M.J., Biomarkers for prostate cancer: prostate-specific antigen and beyond. Clin Chem Lab Med, 2020. 58(3): p. 326–339.
Pinsky, P.F., P.C. Prorok and B.S. Kramer, Prostate Cancer Screening - A Perspective on the Current State of the Evidence. N Engl J Med, 2017. 376(13): p. 1285–1289.
Martin, N.E., et al., Prognostic determinants in prostate cancer. Cancer J, 2011. 17(6): p. 429–37.
Grozescu, T. and F. Popa, Prostate cancer between prognosis and adequate/proper therapy. J Med Life, 2017. 10(1): p. 5–12.
Choi, R.Y., et al., Introduction to Machine Learning, Neural Networks, and Deep Learning. Transl Vis Sci Technol, 2020. 9(2): p. 14.
Karthick K, Aruna SK, Samikannu R, Kuppusamy R, Teekaraman Y, Thelkar AR. Implementation of a Heart Disease Risk Prediction Model Using Machine Learning. Comput Math Methods Med. 2022;2022:6517716.
Lee, C.S. and A.Y. Lee, Clinical applications of continual learning machine learning. Lancet Digit Health, 2020. 2(6): p. e279-e281.
Fizazi, K., et al., Abiraterone plus prednisone added to androgen deprivation therapy and docetaxel in de novo metastatic castration-sensitive prostate cancer (PEACE-1): a multicentre, open-label, randomised, phase 3 study with a 2 x 2 factorial design. Lancet, 2022. 399(10336): p. 1695–1707.
Gnip P, Vokorokos L, Drotár P. Selective oversampling approach for strongly imbalanced data. PeerJ Comput Sci. 2021;7:e604.
Ukey N, Yang Z, Li B, Zhang G, Hu Y, Zhang W. Survey on Exact kNN Queries over High-Dimensional Data Space. Sensors (Basel). 2023;23(2):629.
Langarizadeh, M. and F. Moghbeli, Applying Naive Bayesian Networks to Disease Prediction: a Systematic Review. Acta Inform Med, 2016. 24(5): p. 364–369.
Rigatti, S.J., Random Forest. J Insur Med, 2017. 47(1): p. 31–39.
T. Chen, C. Guestrin, Xgboost: A scalable tree boosting system, CoRR abs/1603.02754. arXiv:1603.02754.
Zheng H, Xiao F, Sun S, Qin Y. Brillouin Frequency Shift Extraction Based on AdaBoost Algorithm. Sensors (Basel). 2022;22(9):3354.
Ishwaran, H. and R. O'Brien, REPLY: THE STANDARDIZATION AND AUTOMATION OF MACHINE LEARNING FOR BIOMEDICAL DATA. J Thorac Cardiovasc Surg, 2022. 163(1): p. e102-e103.
Deo, R.C., Machine Learning in Medicine: Will This Time Be Different? Circulation, 2020. 142(16): p. 1521–1523.
Nayan, M., et al., A machine learning approach to predict progression on active surveillance for prostate cancer. Urol Oncol, 2022. 40(4): p. 161.e1-161.e7.
Lin YT, Lee MT, Huang YC, Liu CK, Li YT, Chen M. Prediction of Recurrence-associated Death from Localized Prostate Cancer with a Charlson Comorbidity Index-reinforced Machine Learning Model. Open Med (Wars). 2019;14:593–606.
Lee, C., et al., Application of a novel machine learning framework for predicting non-metastatic prostate cancer-specific mortality in men using the Surveillance, Epidemiology, and End Results (SEER) database. Lancet Digit Health, 2021. 3(3): p. e158-e165.
Cascardo F, Anselmino N, Páez A, Labanca E, Sanchis P, Antico-Arciuch V, Navone N, Gueron G, Vázquez E, Cotignola J. HO-1 Modulates Aerobic Glycolysis through LDH in Prostate Cancer Cells. Antioxidants (Basel). 2021;10(6):966.
Wei RJ, Li TY, Yang XC, Jia N, Yang XL, Song HB. Serum levels of PSA, ALP, ICTP, and BSP in prostate cancer patients and the significance of ROC curve in the diagnosis of prostate cancer bone metastases. Genet Mol Res. 2016;15(2).

Table 1 Statistical description of baseline data for patients with and without progression

Baseline data	Progress group（n=178）	un-progress group（n=355）	c²/F/N	p-value
age	68±7.9	69±6.9		0.127
Symptoms of obstruction appear			13.14	<0.001
appear	116	284
Not yet appeared	62	71
Symptoms of hematuria appear			0.004	0.95
appear	16	34
Not yet appeared	162	321
Whether nodules are touched			3.383	0.07
touched	106	180
Not touched	72	175
Pathological type			1.151	0.283
Adenocarcinoma	177	347
Non-adenocarcinoma	1	8
Gleason score			18.115	2.079
＜8	38	143
≥8	140	212
Prostatic volume(cm)	117.76±102.65	114.59±106.73	0.327	0.843
fPSA(ng/ml)	24.2 （10.07~50）	14.635 （3.5175~50）	9.671	4.017
Initial TPSA(ng/ml)	474.5 （112.575~1486.75）	206 （50.98~673）	10.007	1.419
First diagnosis of bone metastasis			112.111	3.378
appear	160	147
Not yet appeared	18	208
Initial diagnosis of visceral metastasis			1.239	0.266
appear	20	28
Not yet appeared	158	327
Tumor burden			139.79	2.956
High load	157	119
Low load	21	236
T-stage			47.253	6.238
T1-T2	20	145
T3-T4	158	210
N-stage			67.367	2.254
N0	41	217
N1	137	138
M-stage			107.651	3.206
M0	15	196
M1	163	157
LDH(U/L)	244.39±118.08	195.05±66.47	6.197	1.157
Alkaline phosphatase( U/L)	166.5 （98.25~376）	88 （71~121）	9.027	1.757
Testosterone levels at first diagnosis(ng/dL)	419.31±220.57	423.74±185.54	-0.425	0.671
Whether to remove the prostate			84.767	3.357
Removed	5	147
Unremoved	173	208

No competing interests reported.

Appendix20230831233139.pdf
Figure(S1):Data set of patient age, prostate volume, and distribution of biochemical markers.(a)Age distribution of patients.(b) distribution of Gleason score and WHO group.(c) Prostate volume size distribution of patients in the dataset.(d) fPSA distribution of patients.(e) TPSA distribution of patients.(f) Testosterone distribution at first diagnosis. (g) Distribution of alkaline phosphatase in patients.(h) lactate dehydrogenase distribution of patients in the dataset;Figure (S2) Selection graph of K value in KNN model under grid search;Figure (S3) Adjustment chart of random forest parameters:(a)Random forest max_depth parameter adjustment diagram.(b)Random forest min_samples_leaf parameter adjustment diagram.（c）Random forest min_samples_split parameter adjustment chart.（d）Random forest max_features parameter adjustment diagram.Figure (S4) Diagram of a random forest model for one patient.Figure (S5) Relationship between XGboost max_depth and model accuracy.Figure (S6) XGboost model visualization.Table(S1) Feature subsets of different patient datasets from January 2017 to February 2023.
prostatecancer.xlsx

Application of machine learning methods to predict progression in patients with hormone-sensitive prostate cancer

Status:

Version 1

Abstract

Objective

Methods

Results

Conclusion

Figures

Introduction

Method

Results

Discussion

Conclusion

Abbreviations

Declarations

References

Table

Additional Declarations

Supplementary Files

Status:

Version 1