Machine-learning algorithms for the prediction of adverse prognosis in patients undergoing peritoneal dialysis

doi:10.21203/rs.3.rs-3333476/v1

Download PDF

Research Article

Machine-learning algorithms for the prediction of adverse prognosis in patients undergoing peritoneal dialysis

https://doi.org/10.21203/rs.3.rs-3333476/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 02 Jan, 2024

Read the published version in BMC Medical Informatics and Decision Making →

You are reading this latest preprint version

Background

An appropriate prediction model for the adverse prognosis before peritoneal dialysis (PD) is lacking. Therefore, we retrospectively analyzed patients who received PD to construct a predictive model for adverse prognoses using machine learning (ML).

Methods

A retrospective analysis was conducted on 873 patients who underwent PD from August 2007 to December 2020. Five commonly used machine learning algorithms are used for initial model training. Using the area under the curve and accuracy, we ranked the indicators with the highest impact and displayed them using the Shapley additive explanation (version 0.41.0) values, from which the top 20 indicators were selected to build a compact model conducive to clinical application. All model building steps are implemented in Python (version 3.8.3).

Results

A total of 824 patients were included in the analysis at the end of follow-up, 353 patients withdrew from PD (converted to haemodialysis or died), and 471 patients continued receiving PD. In complete model, the CatBoost model exhibited the strongest performance (AUC: 0.80, 95% CI: 0.76–0.83; ACC: 0.78, 95%CI: 0.72–0.83) and was selected for subsequent analysis. We reconstructed a compression model by extracting 20 key features ranked by the SHAP values, the Catboost model also showed the strongest performance (AUC: 0.79; ACC: 0.74).

Conclusions

The Catboost model built using the intelligent analysis technology of ML demonstrated the best predictive performance. Thus, our developed prediction model has potential value in patient screening before PD and hierarchical management after peritoneal dialysis.

machine learning

peritoneal dialysis

prediction model

prognosis

Peritoneal dialysis (PD) is one of the main renal replacement treatments for end-stage renal disease (ESRD, also known as uraemia)^[1]. The average number of new ESRD diagnoses worldwide is 144 individuals per million of the general population^[2], of which approximately 11% receive PD ^[3]. The international PD guidelines recommend family-based renal replacement therapy owing to the prevalence of coronavirus 2019, and it has become the first choice for dialysis patients because of its simplicity and low cost^[4]. However, factors such as peritonitis and peritoneal fibrosis lead to the failure of PD technology. The failure of PD technology limits its application leading to patient withdrawal owing to the cost, and lowers patient survival rates, even leading to death^{[5, 6]}. Early prediction may screen out patients at high risk of PD technology failure in the short term, and help to decide whether to choose PD for kidney replacement.

With the exponential growth in healthcare data, machine learning (ML) is expected to provide more accurate and personalised services when processing large-scale medical data, predicting the development and prognosis of diseases, assisting doctors in formulating treatment plans, and identifying new disease risk factors and treatment methods. ML is also expected to promote the progress and development of medical science. ML algorithm is used to evaluate the accuracy of predicting cardiovascular events in asymptomatic populations by comparing random survival forests (a machine learning technique) with standard cardiovascular risk scores^[7]. The prognostic factors affecting kidney transplant surgery cover multiple fields of surgery, immunology, epidemiology, and physiology; the large amount of data that is generated can precisely leverage the computational power of ML^[8]. However, studies using ML algorithms for PD-related prognosis are limited.

ML technology was used to predict the prognosis, survival, and death risk factors of PD patients and reported that deep neural networks demonstrated the best predictive performance (AUC: 0.841) ^[9]. In patients with PD-associated peritonitis, traditional microbiology and molecular biology methods are considerably slow and have limited clinical applications. ML has demonstrated the power of using nonlinear methods to mine complex biomedical datasets to rapidly predict the fine reactivity and specificity of the human immune system and target antibiotic medication for early patient treatment^[10]. Myopenia is associated with cardiovascular risk and mortality in patients with PD, and the ML model can effectively predict PD myopenia using simple clinical indicators^[11].

However, we lack an appropriate prediction model for the adverse prognosis before PD; therefore, we constructed a prediction model for the adverse prognosis of PD using ML based on the data from our medical centre.

Subjects

We retrospectively analyzed 873 patients who received PD at our institution, August 2007 and December 2020. The inclusion criteria were: 1) The diagnosis of chronic renal failure and regular PD treatment for over one month. 2) Aged 16 years or over. The exclusion criteria were as follows: 1) Patients with acute kidney injuries, patients who received emergency PD, and patients in renal function recovery. 2) Patients with incomplete baseline data. 3) Patients who received kidney transplantation during follow-up. 4) Patients who stopped communicating with our medical centre. Based on these criteria, 824 patients were included in our subsequent analyses. This study was approved by the Medical Ethics Committee of the Daping Hospital (YYLS2022-210) and informed consent was obtained from all patients.

Demographic and clinical information

All baseline data were collected before PD. Patient demographic data were as follows: age (years), sex (male/female), height (cm), weight (kg), body surface area (m²), body mass index (kg/m²), marital status (unmarried, married, divorced, widowed), education level (primary school and below, junior high school, high school, college, undergraduate, master's degree or above), ethnicity (Han, other ethnic minorities), smoking history (yes/no), history of alcohol consumption (yes/no), systolic blood pressure (mmHg), diastolic blood pressure (SBP)(mmHg), heart rate (beats/min), urine volume (ml/24 h), primary disease, comorbidities, dialysis term (months), previous history of renal replacement therapy (including haemodialysis and kidney transplantation), and medication history. The laboratory data were as follows: haemoglobin (HGB), ferritin, serum iron, serum total iron binding capacity, transferrin saturation, blood calcium, blood phosphorus, intact parathyroid hormone (iPTH), calcium-phosphorus product, alkaline phosphatase, serum albumin, prealbumin, blood sodium, blood potassium, blood chlorine, carbon dioxide binding capacity, creatinine, urea nitrogen, uric acid β2 microglobulin, estimated glomerular filtration rate (eGFR), total cholesterol, triglyceride, low-density cholesterol (LDL), high-density cholesterol (HDL), fasting blood glucose, glycosylated haemoglobin, B-type natriuretic peptide, troponin, creatine kinase (CK), myoglobin (MB), C-reactive protein, erythrocyte sedimentation rate (ESR), hepatitis B surface antigen, hepatitis C antigen/antibody, syphilis antibody, and human immunodeficiency virus (HIV) antibody.

The imaging techniques were as follows: echocardiography (left ventricular end-diastolic diameter, interventricular septum thickness, left ventricular posterior wall thickness, and calculation of the left ventricular mass and left ventricular mass index), and carotid artery colour Doppler ultrasound (the presence of plaque formation).

The adverse prognosis was defined as the withdrawal from PD or all-cause mortality within 24 months of PD initiation. The patients were divided into PD withdrawal and PD continuation groups according to whether an adverse prognostic event occurred. If a patient withdrew from PD during follow-up, the time and reason for withdrawal (peritonitis, insufficient dialysis, ultrafiltration failure, thoracoabdominal fistula, catheter dysfunction, patient requirements, and other causes) were recorded. If a patient died, the dialysis duration and cause of death (cardiovascular death, other causes) were recorded. The study was terminated on 31 December 2020.

Statistical analysis

The measurement data are expressed as the mean ± standard deviation, and the counting data are expressed as a percentage. The measurement data between the groups were compared using the t-test, and the two group rates were compared using the chi-square test. The data were processed using the Statistical Package for the Social Sciences version 20.0 software. ML methods were used to construct a predictive model for the adverse prognosis in PD patients. During the model construction, the enrolled patients were randomly divided into two groups at a ratio of 7:3. The larger group was the training subset for ML, and the smaller group was the testing subset for model testing. A small number of missing continuous variables were supplemented using the median method, and the categorical variables were supplemented using the 0-value method.

The following five commonly used ML algorithms were used for the initial model training: Cat Boost Classifier (CatBoost, version 1.0.6), Logistic Regression (LR, version 1.0.2), Light Gradient Boosting (LGB, version 3.2.1), Gradient Boosting (GBT, version 1.0.2), and Random Forest (RFL, version 1.0.2). LR is a kind of generalized linear regression. The advantage is that the rate function is derivable to any order and has good mathematical properties. Many existing numerical optimization algorithms can be used to find the optimal solution. The disadvantage is that LR can not be used to solve nonlinear problems, and it is difficult to deal with the problem of data imbalance. In ML, the goal is to train the model most successfully with multiple learning algorithms. Boosting is a method often used in practice that is not built in parallel, but sequentially. The weak algorithm first trains the model, and then reassembles the model according to the training results to make the model easier to learn. GBT is the most basic Boosting model algorithm, which has no role in optimizing complex types of data and missing data. LGB is a very effective way to reduce errors and thus improve accuracy and speed. However, it does not support string type data and requires a special algorithm to split the classified data. LGB performs better on large datasets and high-dimensional data, while CatBoost is better at handling category features and missing values. Categorical Boosting (CatBoost) is a symmetric decision tree based learner, which relies on GBT framework with fewer parameters, support categorical variables and high accuracy. The optimization algorithm formula of CatBoost is as follows:

Where p is the added prior term and α is usually a weight coefficient greater than 0. For binary classification problems, the prior term is the prior probability of the positive example.

The area under the curve (AUC) and the accuracy (ACC) were used as the primary evaluation indices to select the optimal model for further optimisation. We obtained the ranking of the indicators that had the most significant impact on the model and displayed them using the Shapley additive explanation (SHAP, version 0.41.0) values, from which the top 20 indicators were selected to build a compact model conducive to clinical application. All the model-building steps were implemented in Python (version 3.8.3).

According to the inclusion criteria, nine patients with a dialysis duration of less than one month at the time of withdrawal were excluded. During follow-up, 31 patients underwent kidney transplantation, and nine were lost. A total of 824 patients were included in the analysis at the end of follow-up, 353 patients withdrew from PD (converted to haemodialysis or died), and 471 patients continued receiving PD.

Our cohort included 481 men and 343 women with an average age of 47.82 ± 15.45 years, and most were married (91.6%) and of the Han ethnicity (94.7%). The education level was mainly middle school (37.0%), with smokers accounting for 24.7% and alcohol consumption accounting for 15.6%. The three primary causes of chronic renal failure were chronic glomerulonephritis in 479 patients (58.1%), diabetic nephropathy in 112 patients (13.5%), and hypertensive renal injury in 21 patients (2.5%). The most common complications were hypertension (n = 454, 55.0%). Significant differences in age, education level, urinary output, history of kidney transplantation, primary renal disease, comorbidity, and history of medication were identified between the two groups. Demographic data are presented in Table 1.

Table 1

Patient demographics and clinical characteristics
Characteristics	Total (N = 824)	PD continuation group (N = 471)	PD withdrawal group (N = 353)	P
Age (years)	47.82 ± 15.45	45.36 ± 14.46	51.11 ± 16.12	< 0.001
Sex (male)/N (%)	481 (58.3%)	263 (55.8%)	218 (61.7%)	0.088
BMI, kg/m2	22.98 ± 3.62	22.89 ± 3.74	23.1 ± 3.46	0.421
Body surface area (m2)	1.61 ± 0.19	1.6 ± 0.19	1.62 ± 0.18	0.35
Marital status				0.216
Unmarried	62 (7.5%)	42 (8.9%)	20 (5.6%)
Married	755 (91.6%)	426 (90.4%)	329 (93.2%)
Divorced	6 (0.7%)	3 (0.6%)	3 (0.8%)
Widowed	1 (0.1%)	0	1 (0.2%)
Education level				0.004
< 6 years	185 (22.4%)	69 (14.6%)	116 (32.8%)
6–9 years	305 (37.0%)	177 (37.5%)	128 (36.2%)
9–12 years	151 (18.3%)	97 (20.8%)	54 (15.2%)
12–17 years	81 (9.8%)	54 (11.4%)	27 (7.6%)
༞17 years	3 (0.3%)	2 (0.4%)	1 (0.2%)
Ethnicity, Han/N (%)	781 (94.7%)	449 (95.3%)	332 (94.0%)	0.414
Smoking history/N (%)	204 (24.7%)	115 (24.4%)	89 (25.2%)	0.793
Drinking History/N (%)	129 (15.6%)	73 (15.4%)	56 (15.8%)	0.887
Systolic blood pressure (mmHg)	150.47 ± 26.13	150.14 ± 25.34	150.92 ± 27.19	0.672
Diastolic blood pressure (mmHg)	86.23 ± 18.79	87.12 ± 18.08	85.03 ± 19.67	0.115
Heart rate (bpm)	86.28 ± 13.86	87.09 ± 14.28	85.19 ± 13.22	0.052
Urinary output (ml/24 h)	1013.71 ± 473.18	1048.09 ± 448.97	967.84 ± 500.65	0.016
History of kidney transplantation/N (%)	8 (0.9%)	1 (0.2%)	7 (1.9%)	0.027
History of haemodialysis/N (%)	36 (4.3%)	17 (3.6%)	19 (5.3%)	0.218
Primary renal disease				0.001
Glomerulonephritis	479 (58.1%)	298 (63.2%)	181 (51.2%)
Diabetic kidney disease	112 (13.5%)	42 (8.9%)	70 (19.8%)
Hypertension	21 (2.5%)	9 (1.9%)	12 (3.3%)
Obstructive nephropathy	10 (1.2%)	6 (1.2%)	4 (1.1%)
Lupus nephritis	8 (0.9%)	3 (0.6%)	5 (1.4%)
Cystic kidney disease	5 (0.6%)	3 (0.6%)	2 (0.5%)
Renal vasculitis	5 (0.6%)	3 (0.6%)	2 (0.5%)
Others	28 (3.3%)	13 (2.7%)	15 (4.2%)
Unknown	156 (18.9%)	94 (19.9%)	62 (17.5%)
Comorbidity/N (%)
Diabetes mellitus	135 (16.3%)	58 (12.3%)	77 (21.8%)	< 0.001
Hypertension	454 (55.0%)	257 (54.5%)	197 (55.8%)	0.723
Coronary heart disease Or myocardial infarction	48 (5.8%)	25 (5.3%)	23 (6.5%)	0.464
Congestive heart failure	17 (2.0%)	7 (1.4%)	10 (2.8%)	0.178
Cardiac arrhythmias	14 (1.6%)	9 (1.9%)	5 (1.4%)	0.587
History of stroke or Cerebral vascular diseases	25 (3.0%)	8 (1.6%)	17 (4.8%)	0.01
Malignancies	7 (0.8%)	3 (0.6%)	4 (1.1%)	0.442
Peripheral arterial disease	8 (0.9%)	3 (0.6%)	5 (1.4%)	0.259
Urology procedures	12 (1.4%)	6 (1.2%)	5 (1.6%)	0.86
History of medication
ARB	376 (45.6%)	220 (46.7%)	156 (44.1%)	0.473
ACEI	40 (4.8%)	11 (2.3%)	29 (8.2%)	< 0.001
CCB	654 (79.3%)	369 (78.3%)	285 (80.7%)	0.401
Diuretic	160 (19.4%)	99 (21.0%)	61 (17.2%)	0.179
EPO	604 (73.3%)	368 (78.1%)	236 (66.8%)	< 0.001
Uric acid-lowering Medications	110 (13.3%)	87 (18.4%)	23 (6.5%)	< 0.001
Iron	144 (17.4%)	74 (15.7%)	70 (19.8%)	0.123
β-receptor blockade	185 (22.4%)	109 (23.1%)	76 (21.5%)	0.583
α-receptor blockade	187 (22.6%)	95 (20.1%)	92 (26.0%)	0.046
α/β-receptor blockade	60 (7.2%)	25 (5.3%)	35 (9.9%)	0.012
α-ketoacids	226 (27.4%)	112 (23.7%)	114 (32.2%)	0.007
Antidiabetic agents	115 (13.9%)	49 (10.4%)	66 (18.6%)	0.001
Lipid-lowering medications	115 (13.9%)	68 (14.4%)	42 (11.8%)	0.289
Sleep aids	11 (1.3%)	6 (1.2%)	5 (1.4%)	0.109
Glucocorticoids	16 (1.9%)	6 (1.2%)	10 (2.8%)	0.154
Immunosuppressive agents	10 (1.2%)	3 (0.6%)	7 (1.9%)	0.081
Calcimimetic agents	3 (0.3%)	2 (0.4%)	1 (0.2%)	1
BMI = body mass index; PD = peritoneal dialysis; ARB; angiotensin receptor blockers; ACEI, angiotensin-converting enzyme inhibitors; CCB, calcium channel blockers; EPO, erythropoietin.

The PD withdrawal group had higher levels of ferritin, blood calcium, alkaline phosphatase, estimated glomerular filtration rate (eGFR), low-density cholesterol, fasting blood glucose, glycated haemoglobin, C-reactive protein, cardiac ejection fraction, blood phosphorus, iPTH, serum albumin, creatinine, urea nitrogen, and uric acid. β2 microglobulin was lower in the PD continuation group, and we observed no statistically significant differences in the other indicators (Table 2).

Table 2

Patient baseline laboratory data for peritoneal dialysis
Characteristics	Total (N = 824)	PD continuation group (N = 471)	PD withdrawal group (N = 353)	P
Haptoglobin	81.23 ± 18.41	81.02 ± 19.05	81.5 ± 17.55	−0.376
Ferritin	290.04 ± 289.57	270.31 ± 263.08	316.78 ± 320.55	−2.06
Serum iron	12.67 ± 7.58	12.72 ± 7.61	12.6 ± 7.55	0.21
Total iron binding capacity	44.17 ± 9.19	43.84 ± 8.82	44.66 ± 9.73	−1.189
Transferrin saturation	29.35 ± 17.16	29.69 ± 16.89	28.85 ± 17.58	0.651
Serum calcium	1.93 ± 0.28	1.91 ± 0.29	1.95 ± 0.27	−2.188
Serum phosphorus	2.03 ± 0.62	2.1 ± 0.62	1.94 ± 0.6	3.699
Intact parathormone	396.61 ± 257.89	420.99 ± 264.14	359.5 ± 243.89	3.19
Calcium-phosphorus product	4.15 ± 1.24	4.2 ± 1.2	4.09 ± 1.29	1.243
Alkaline phosphatase	92.63 ± 45.68	89.52 ± 42.86	97.27 ± 49.33	−2.149
Serum albumin	32.84 ± 5.61	33.79 ± 5.21	31.58 ± 5.88	5.707
Prealbumin	302.51 ± 88.02	309.74 ± 82.38	291.27 ± 95.2	2.692
Serum sodium	138.31 ± 4.01	138.5 ± 3.95	138.05 ± 4.07	1.615
Serum potassium	4.59 ± 0.87	4.59 ± 0.82	4.59 ± 0.94	−0.028
Serum chloremia	104.88 ± 5.81	105.21 ± 5.74	104.44 ± 5.89	1.879
Carbon dioxide combining power	18.84 ± 4.67	18.72 ± 4.65	19.01 ± 4.7	−0.882
Serum creatinine	891.45 ± 358.22	920.84 ± 359.19	852.24 ± 353.64	2.731
Serum urea nitrogen	28.48 ± 12.96	29.85 ± 13.92	26.66 ± 11.34	3.513
Uric acid	492.67 ± 139.6	502.99 ± 136.61	478.93 ± 142.52	2.455
β2 microglobulin	21.55 ± 8.74	22.67 ± 8.58	19.89 ± 8.72	3.669
eGFR	5.68 ± 2.31	5.47 ± 2.13	5.96 ± 2.5	−3.036
Total cholesterol	4.15 ± 1.24	4.13 ± 1.25	4.18 ± 1.23	−0.571
Triglyceride	1.54 ± 1.07	1.58 ± 1.08	1.49 ± 1.04	1.057
Lower blood lipids	2.52 ± 0.88	2.59 ± 0.87	2.43 ± 0.89	2.48
High-density lipoprotein	1.17 ± 0.52	1.16 ± 0.49	1.2 ± 0.56	−0.942
Fasting blood glucose	5.07 ± 1.97	4.88 ± 1.46	5.33 ± 2.49	−2.91
Glycosylated haemoglobin A1c	5.33 ± 0.9	5.21 ± 0.88	5.5 ± 0.9	−3.168
B-type Natriuretic Peptide	2062.5 ± 4874.21	1834.29 ± 4537.45	2500.66 ± 5485.42	−0.783
Troponin	0.07 ± 0.09	0.07 ± 0.08	0.08 ± 0.1	−1.245
Creatine kinase	3.67 ± 6.79	3.57 ± 6.97	3.86 ± 6.44	−0.466
Myoglobin	279.14 ± 342.13	279.44 ± 337.61	278.58 ± 351.5	0.028
C-reactive protein	12.36 ± 27.05	9.25 ± 21.24	17.67 ± 34.2	−3.367
Erythrocyte sedimentation rate	56.2 ± 38.08	54.69 ± 35.74	59.29 ± 42.46	−0.994
Hepatitis B surface antigen	70 (8.4%)	40 (8.4%)	30 (8.4%)	< 0.001
Hepatitis C antigen	11 (1.3%)	4 (0.8%)	7 (1.9%)	1.202
Syphilis antibody	2 (0.2%)	1 (0.2%)	1 (0.2%)	< 0.001
Left ventricular end-diastolic dimensions	4.93 ± 3.76	4.81 ± 3.01	5.26 ± 5.38	−0.879
Interventricular septal thickness	1.19 ± 0.76	1.16 ± 0.6	1.29 ± 1.11	−1.256
left ventricular posterior wall thickness	1.07 ± 0.16	1.07 ± 0.16	1.08 ± 0.17	−0.516
left ventricular mass	217.97 ± 70.66	215.91 ± 70.97	226.39 ± 69.49	−0.921
left ventricular mass index	133.33 ± 40.04	132.18 ± 40.18	138.03 ± 39.53	−0.906
left ventricular ejection fraction	64.02 ± 7.07	64.87 ± 6.63	61.67 ± 7.76	3.331
Carotid artery plaque/N (%)	28 (3.3%)	16 (3.3%)	12 (3.3%)	< 0.001
eGFR = estimated glomerular filtration rate; PD = peritoneal dialysis

Comparison of the five complete models

The performances of the different models are listed in Table 3. The CatBoost algorithm exhibited excellent AUC (0.80, 95% CI: 0.76–0.83) and ACC (0.78, 95% CI: 0.72–0.83) values. The prediction performance of the traditional LR method was acceptable, with an AUC of 0.76 (95% CI: 0.73–0.80) and an ACC of 0.71 (95% CI: 0.64–0.77). The performances of the other three ensemble learning algorithms, LGB (AUC: 0.72; ACC: 0.74), GBT (AUC: 0.72; ACC: 0.76), and RFL (AUC: 0.72; ACC: 0.65), were relatively poor. The ROC curve of the complete model is displayed in Fig. 1a. The CatBoost model exhibited the strongest performance and was selected for subsequent analysis.

Table 3

Basic performance indicators of the five complete models
Model name	AUC (CI)	ACC (CI)
Cat Boost Classifier	0.8	0.78
Cat Boost Classifier	[0.76, 0.83]	[0.72, 0.83]
Logistic Regression	0.76	0.71
Logistic Regression	[0.73, 0.80]	[0.64, 0.77]
Light Gradient Boosting	0.72	0.74
Light Gradient Boosting	[0.68, 0.77]	[0.68, 0.81]
Gradient Boosting	0.72	0.76
Gradient Boosting	[0.70, 0.79]	[0.70, 0.82]
Random Forest	0.72	0.65
Random Forest	[0.62, 0.75]	[0.58, 0.72]
ACC = accuracy; AUC = area under the curve; CI = confidence interval

Key features and compact model

After eliminating 187 individuals due to missing data on covariates or the predictor variables of interest, 637 patients were included in the final model construction. The calculated SHAP values summarised the ranking of the features that had the strongest influence on the prediction results of the complete model. The feature names and the extent of their influence are presented in Fig. 2. Among the demographic characteristics, age, weight, BMI, and education level were selected as significant predictors of adverse PD prognosis. Iron metabolism was closely related to haematopoiesis, among which TIBC and SF were key predictors. Prealbumin and serum albumin levels, which were closely related to liver synthesis, also played important roles. There were also prominent roles for HDL cholesterol, FBG, and total cholesterol in glycolipid metabolism. In addition, calcium and phosphorus metabolism (Vitamin D, serum phosphorus, and iPTH), cardiovascular function (SBP, CK, MB), ESR, and creatinine had some predictive effects.

We reconstructed a compression model by extracting 20 key features ranked by the SHAP values. This simplified version of the model (AUC: 0.79; ACC: 0.74) was slightly weaker in performance than the full model but was more conducive to clinical application and data collection (Fig. 1b). The complement model had a maximum Youden index of 0.48, which gives a sensitivity of 0.68 and a specificity of 0.80. The maximum Youden index of the compact model was 0.46, and the sensitivity and specificity were 0.71 and 0.75, respectively. The specific data parameters are listed in Table 4.

Table 4

Performance indicators of the final models
Model	Performance
Model	ACC	AUC	Youden	Sensitivity	Specificity
Full	0.78 [0.72, 0.83]	0.80 [0.76, 0.83]	0.48	0.68 [0.53, 0.80]	0.80 [0.73, 0.86]
Compact	0.74 [0.68, 0.80]	0.79 [0.75, 0.84]	0.46	0.71 [0.56, 0.82]	0.75 [0.67, 0.81]
ACC = accuracy; AUC = area under the curve

Model explanation

The summary plot of the SHAP values in Fig. 2 provides an overview of the impact of the features of the final model. Figure 3 illustrates two specific forecasting examples. The blue bars represent protective factors, where longer bars indicate that PD was less likely to fail. The red bars represent risk factors and indicate the opposite effect. As depicted in Fig. 3a, a 64-year-old patient with poor education, low TIBC, and prealbumin levels was suspected of having reduced hepatic compensatory function. The final model predicted that his PD would fail, and he became hospitalised with infectious peritonitis after 7.1 months of dialysis. For another 40-year-old patient, HGB, prealbumin, and albumin levels all appeared normal, indicating a strong compensatory capacity (Fig. 3b). The model predicted that the patient was suitable for PD, and the patient continued PD after the follow-up period for over five years.

PD-associated peritonitis is one of the leading causes of PD withdrawal and death.^{[12, 13]} ML algorithms are becoming increasingly popular in medical research and can be applied to disease screening, diagnosis, and prognosis. We used ML intelligent analysis technology to construct a predictive model for the adverse prognosis of PD and demonstrated that age, body weight, and albumin levels are important predictive factors for the adverse prognosis of PD. We developed five predictive models; in the complete model, the calculated SHAP values summarised the strongest predictive indicators and sorted and extracted the 20 key features to reconstruct the model. Collectively, our findings suggested that the CatBoost model demonstrated the strongest performance.

We ranked the factors closely related to the adverse prognosis of patients by the SHAP values, with the top 20 key factors including age, body weight, albumin, and blood lipids. The meta-analysis revealed that age is a risk factor for all-cause cardiovascular death in dialysis patients^[14]. In this study, we observed that the age of patients in the PD continuation group was significantly lower than that in the adverse prognosis group (45.36 vs 51.11 years, P < 0.001). In the complete model, the calculated SHAP values confirmed that age had the strongest impact on predicting an adverse prognosis for PD patients. In addition, body weight and BMI were critical predictive factors for adverse PD prognosis, with higher BMI leading to higher hospitalization rates for peritonitis^[15]. In the general population, obesity is associated with increased cardiovascular risk and reduced survival, but the “obesity paradox” in ESRD has always been controversial^{[16, 17]}. Our study suggests that increased body weight and BMI correlate with a lower risk of adverse PD prognosis. The nutritional indicators include body weight, as well as albumin and blood lipids. A positive correlation between nutritional status and dialysis duration has been reported in PD patients because a nutritious diet reduces the incidence of complications such as peritonitis^[12].

Education level was also considered a vital predictor of adverse PD prognosis, and multiple studies have demonstrated that ^{[18, 19]} patients with lower education levels experience increased peritonitis and technical failure than those with higher education levels. The potential reason may be that patients with lower education levels have lower incomes and poor compliance, which affects their access to timely healthcare, medication, and treatment.

The high prevalence of cardiovascular diseases in PD patients is related to uremic toxins, inflammation (erythrocyte sedimentation rate), and disorders in bone mineral metabolism (vitamin D, serum phosphorus, and iPTH)^[20]. Similarly, we observed vitamin D, serum phosphorus, iPTH, erythrocyte sedimentation rate, creatinine, and cardiovascular disease to be associated with adverse PD prognosis in patients. Furthermore, we observed that TIBC and SF are critical predictive factors for adverse PD prognosis and that higher amounts of iron increase the risk of QT dispersion^[21]. Functional iron deficiency is an independent risk factor for all-cause death in PD patients. Consistent with our research findings, PD patients with high iron levels have a four-fold higher risk of all-cause cardiovascular death ^[22].

ML is an interdisciplinary field of mathematics and statistics^[23] that involves fitting predictive models to data for information grouping. We assumed that ML methods could predict the adverse prognosis of patients before starting PD, recommended the most favourable dialysis method, and provided timely medical intervention, which improved patient prognosis and reduced medical costs.

CatBoost is the third GBDT-based improved algorithm after XGBoost and LightGBM^[24]. Launched by Yandex Company in Russia in 2018, and is open-source. It uses gradient lifting on the decision tree and can be easily integrated into deep-learning frameworks. Based on the GBDT framework, which has fewer parameters, CatBoost supports categorical variables with high accuracy and can efficiently and reasonably process t- algorithms. CatBoost has been extensively studied in the prediction of skin sensitisation ^[25], depression occurrence^[26], pregnancy diabetes management^[27], and transplanted kidney function^[8], and it exhibits good predictive performance. Owing to numerous factors that affect an adverse PD prognosis and considering the clinical applications, we reconstructed a compression model by extracting 20 key features ranked by the SHAP values. This simplified version of the model was slightly weaker in performance than the full model but was more conducive to clinical application and data collection. Before a patient starts PD, the CatBoost model can be used to predict whether the patient is suitable for PD treatment and whether PD-related peritonitis may occur. Based on the prediction, the most optimal dialysis plan can be selected for the patient allowing early intervention.

Our study had several limitations. First, this was a single-centre retrospective study, and we could not evaluate whether the external cohort population exhibited the same pattern. Second, this study used the median of missing values, which inevitably led to bias. Third, the number of cases was relatively small, and the model construction lacked cross-validation and external validation, all of which affected the ability to generalise the model. A multicenter joint study is needed to validate the model.

Collectively, the Catboost model built using the intelligent analysis technology of ML demonstrated the best predictive performance (AUC: 0.79; ACC: 0.74). Thus, the model has potential value in patient screening before PD and hierarchical management after peritoneal dialysis.

Peritoneal dialysis: PD

Machine learning: ML

End-stage renal disease: ESRD

Haemoglobin: HGB

Intact parathyroid hormone: iPTH

Estimated glomerular filtration rate: eGFR

Low-density cholesterol: LDL

High-density cholesterol: HDL

Creatine kinase: CK

Myoglobin: MB

C-reactive protein: CRP

Erythrocyte sedimentation rate: ESR

Human immunodeficiency virus: HIV

Logistic Regression: LR

Light Gradient Boosting: LGB

Gradient Boosting: GBT

Random Forest: RFL

Area under the curve: AUC

Accuracy: ACC

Shapley additive explanation: SHAP

Acknowledgements

The authors would like to thank Professor Jie Yang of the Daping Hospital for providing statistical support.

Authors' contributions

JY, JF, and LF are the joint ﬁrst authors. KC, LX and JY designed the study. JY, JF and LF performed the experiments. JY, SH, and KY collected the data. LF, LF, and LX analyzed the data. JF, LX, and KC wrote the manuscript. All of the authors have read and approved the final version of the manuscript.

Funding

This work was supported by grants from the National Natural Science Foundation of China (82270768), the Chongqing Technology Innovation project (2022YSZX- JCX0007CSTB, cstc2019jscx- msxmX0258), and the National Science and Technology Support Plan (2022-173ZD-112, SKLKF202201).

Availability of data and materials

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Ethics approval and consent to participate

This study was approved by the Medical Ethics Committee of the Daping Hospital (YYLS2022-210), prior to the commencement of this study, and the requirement for informed consent was waived, as the utilization of anonymized retrospective data does not require patient consent under the local legislation. In addition to this, all methods were carried out following relevant guidelines and regulations.

Competing interests.

The authors declare that they have no competing interests.

Shrestha BM. Peritoneal Dialysis or Haemodialysis for Kidney Failure?[J]. JNMA J Nepal Med Assoc 2018,56(210):556–7.
Thurlow JS, Joshi M, Yan G, et al. Global Epidemiology of End-Stage Kidney Disease and Disparities in Kidney Replacement Therapy[J]. Am J Nephrol. 2021;52(2):98–107.
Mehrotra R, Devuyst O, Davies SJ, et al. The Current State of Peritoneal Dialysis[J]. J Am Soc Nephrol. 2016;27(11):3238–52.
Teitelbaum I. Peritoneal Dialysis[J]. N Engl J Med. 2021;385(19):1786–95.
Chaudhary K, Sangha H, Khanna R. Peritoneal dialysis first: rationale[J]. Clin J Am Soc Nephrol. 2011;6(2):447–56.
Global regional, national burden of chronic kidney disease. and, 1990–2017: a systematic analysis for the Global Burden of Disease Study 2017[J]. Lancet,2020,395(10225):709–33.
Venkatesh R, Balasubramanian C, Kaliappan M. Development of Big Data Predictive Analytics Model for Disease Prediction using Machine learning Technique[J]. J Med Syst. 2019;43(8):272.
Quinino RM, Agena F, Modelli DAL et al. A Machine Learning Prediction Model for Immediate Graft Function After Deceased Donor Kidney Transplantation[J]. Transplantation,2023.
Noh J, Yoo KD, Bae W, et al. Prediction of the Mortality Risk in Peritoneal Dialysis Patients using Machine Learning Models: A Nation-wide Prospective Cohort in Korea[J]. Sci Rep. 2020;10(1):7470.
Zhang J, Friberg IM, Kift-Morgan A, et al. Machine-learning algorithms define pathogen-specific local immune fingerprints in peritoneal dialysis patients with bacterial infections[J]. Kidney Int. 2017;92(1):179–91.
Wu J, Lin S, Guan J et al. Prediction of the sarcopenia in peritoneal dialysis using simple clinical information: A machine learning-based model[J]. Semin Dial,2023.
Kiebalo T, Holotka J, Habura I et al. Nutritional Status in Peritoneal Dialysis: Nutritional Guidelines, Adequacy and the Management of Malnutrition[J]. Nutrients,2020,12(6).
Cho Y, Johnson DW. Peritoneal dialysis-related peritonitis: towards improving evidence, practices, and outcomes[J]. Am J Kidney Dis 2014,64(2):278–89.
Zhang J, Lu X, Li H, et al. Risk factors for mortality in patients undergoing peritoneal dialysis: a systematic review and meta-analysis[J]. Ren Fail. 2021;43(1):743–53.
Obi Y, Streja E, Mehrotra R, et al. Impact of Obesity on Modality Longevity, Residual Kidney Function, Peritonitis, and Survival Among Incident Peritoneal Dialysis Patients[J]. Am J Kidney Dis. 2018;71(6):802–13.
Park J, Ahmadi SF, Streja E et al. Obesity paradox in end-stage kidney disease patients[J]. Prog Cardiovasc Dis 2014,56(4):415–25.
Prasad N, Sinha A, Gupta A et al. Effect of body mass index on outcomes of peritoneal dialysis patients in India[J]. Perit Dial Int 2014,34(4):399–408.
Kim HJ, Lee J, Park M, et al. Lower Education Level Is a Risk Factor for Peritonitis and Technique Failure but Not a Risk for Overall Mortality in Peritoneal Dialysis under Comprehensive Training System[J]. PLoS ONE. 2017;12(1):e169063.
Fan X, Huang R, Wang J et al. Risk factors for the first episode of peritonitis in Southern Chinese continuous ambulatory peritoneal dialysis patients[J]. PLoS One 2014,9(9):e107485.
Albakr RB, Bargman JM. A Comparison of Hemodialysis and Peritoneal Dialysis in Patients with Cardiovascular Disease[J]. Cardiol Clin,2021,39(3):447–53.
Bavbek N, Yilmaz H, Erdemli HK et al. Correlation between iron stores and QTc dispersion in chronic ambulatory peritoneal dialysis patients[J]. Ren Fail 2014,36(2):187–90.
Luo D, Zhong Z, Qiu Y et al. Abnormal iron status is associated with an increased risk of mortality in patients on peritoneal dialysis[J]. Nutr Metab Cardiovasc Dis 2021,31(4):1148–55.
Greener JG, Kandathil SM, Moffat L, et al. A guide to machine learning for biologists[J]. Nat Rev Mol Cell Biol. 2022;23(1):40–55.
Yandex. CatBoost[Z]. 2018.
Ambe K, Suzuki M, Ashikaga T et al. Development of quantitative model of a local lymph node assay for evaluating skin sensitization potency applying machine learning CatBoost[J]. Regul Toxicol Pharmacol,2021,125:105019.
Zhang C, Chen X, Wang S et al. Using CatBoost algorithm to identify middle-aged and elderly depression, national health and nutrition examination survey 2011–2018[J]. Psychiatry Res,2021,306:114261.
Kumar M, Ang LT, Ho C et al. Machine Learning-Derived Prenatal Predictive Risk Model to Guide Intervention and Prevent the Progression of Gestational Diabetes Mellitus to Type 2 Diabetes: Prediction Model Development Study[J]. JMIR Diabetes,2022,7(3):e32366.

No competing interests reported.

Download PDF

Journal Publication

published 02 Jan, 2024

Read the published version in BMC Medical Informatics and Decision Making →

Editorial decision: Major revision
15 Sep, 2023
Editor assigned by journal
15 Sep, 2023
Submission checks completed at journal
14 Sep, 2023
First submitted to journal
07 Sep, 2023

You are reading this latest preprint version

Machine-learning algorithms for the prediction of adverse prognosis in patients undergoing peritoneal dialysis

Status:

Journal Publication

Version 1

Abstract

Figures

Background

Materials and methods

Subjects

Demographic and clinical information

Statistical analysis

Results

Comparison of the five complete models

Key features and compact model

Model explanation

Discussion

Conclusions

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1