https://orcid.org/0000-0001-5206-0804
Abstract
The billing database of the universal healthcare system in Japan potentially includes large-cohort data of patients with immunoglobulin A nephropathy, diagnosis codes aimed at billing should not be directly used for clinical research because of the risk of misdiagnosis. To solve this problem, we aimed to develop a novel method for identifying patients with immunoglobulin A nephropathy from billing data using machine learning. The medical records and bills of 3,743 patients who consulted nephrologists at a single center were extracted. Patients were labeled to have been diagnosed with immunoglobulin A nephropathy through a review of medical records. A manual analysis of the diagnostic accuracy and machine learning was performed. For machine learning, the datasets were preprocessed in three patterns and assigned to the XGBoost program using five-fold cross-validation. Of all the participants, 437 were labeled as having been diagnosed with immunoglobulin A nephropathy. Bill codes for immunoglobulin A nephropathy were provided to approximately half of them. The manually created criteria consisting of the recommended examinations and treatments in the Japanese guidelines for immunoglobulin A nephropathy showed both specificity and sensitivity < 0.8. In contrast, with the receiver operating characteristic curve analysis, the machine learning process yielded area under the curve values over 0.9 with preprocessing from the clinical viewpoint. Applying machine learning technology to a dataset preprocessed from a clinical viewpoint achieved a high performance in detecting patients with immunoglobulin A nephropathy. This methodology contributes to the construction of a disease-specific cohort using big bill data.
Figures
Citation: Tsunoda R, Kume K, Kagawa R, Sanuki M, Kitagawa H, Mase K, et al. (2024) Machine-learning-based identification of patients with IgA nephropathy using a computerized medical billing database. PLoS ONE 19(12): e0312915. https://doi.org/10.1371/journal.pone.0312915
Editor: V. Vinoth Kumar, Vellore Institute of Technology, INDIA
Received: January 30, 2023; Accepted: October 16, 2024; Published: December 5, 2024
Copyright: © 2024 Tsunoda et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: Data cannot be shared publicly because of Japanese Act on the Protection of Personal Information. Data access requests may be sent to the University of Tsukuba Hospital Director, Yuji Hiramatsu, at hsp.somuka@un.tsukuba.ac.jp.
Funding: This study is supported by funds below. Japan Society for the Promotion of Science KAKENHI, JP19H04114 Japan Society for the Promotion of Science KAKENHI, JP21K08220 Japan Society for the Promotion of Science KAKENHI, JP19K19347 Japan Agency for Medical Research and Development, JP21zf0127005 Japan Science and Technology Agency -Mirai Program, JPMJMI19G8 programs for Progress of the next Cross-ministerial Strategic Innovation Promotion Program (SIP) on “Integrated Health Care System(C-1)”, Cabinet Office, Government of Japan, JPJ012425 The recipient of all funds is Kunihiro Yamagata, M.D.,Ph.D‥ The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Immunoglobulin A (IgA) nephropathy (IgAN) is the most common type of glomerulonephritis [1]. The incidence of IgAN varies by region, with an annual incidence (per 100,000 population) of 0.76–2.5 in Europe [2,3] and 4.5 in Japan [4]; although the estimated point prevalence is 2.53 per 10,000 in Europe [3], it may be higher in Japan. The pathogenesis of IgAN warrants long-term observations, over decades, to evaluate the effects of different treatments on renal outcomes [5,6]. Thus, there is a strong need for large clinical studies to determine the treatment efficacy based on long-term observations.
Under the universal healthcare system, the National Database of Health Insurance Claims and Specific Health Checkups of Japan (NDB) contains the billing data of almost all citizens in Japan [7]. Although this database has been used for research earlier, the majority were descriptive or cross-sectional studies [8] because of the paucity of clinical information on the diseases of patients. Thus, using this approach, it is difficult to estimate the survival of patients.
In contrast, end-stage kidney disease is almost always fully recorded on the bill at the time of initiation of dialysis or kidney transplantation. Thus, identifying patients with specific kidney diseases, such as IgAN, from the billing database might be an innovative method of creating the largest IgAN database that includes the complete implementation record of examinations and treatments. However, a critical barrier to the extraction of specific diagnostic information using billing data is that the diagnoses recorded on medical bills do not necessarily correspond to the patients’ actual clinical conditions because the diagnostic codes have been given only for reward claims. Correct diagnoses cannot be obtained using diagnostic codes alone, and this issue has been persistently reported by researchers [9]. In the field of nephrology, the coding of CKD and acute kidney disease has low sensitivity for a true clinical diagnosis [10]. Moreover, the inaccurate diagnostic codes in billing data have been reported for many diseases, including schizophrenia [11], hypertension [12], diabetes mellitus [13], rheumatological diseases [14], and cardiovascular events [15]. Thus, there exists a serious risk of misdiagnosis when only diagnostic codes are used to identify patients with any specific disease.
There are two types of biases in the diagnostic codes for medical bills. One bias is “underdiagnosis.” In other words, an important diagnosis that does not pertain to the claim for payment is not necessarily included in bills. In most patients who consult a nephrologist for the treatment of IgAN, a diagnosis of “CKD” or “chronic glomerulonephritis” is sufficient for an appropriate claim because there is no test or treatment specific for IgAN. It is difficult to correctly detect patients with IgAN using codes alone.
Another bias is “overdiagnosis.” It is possible that the given diagnostic codes do not fully match the patient’s clinical condition. For example, patients diagnosed with hypertension “hypertension.” In Japan, angiotensin II receptor blockers, the recommended drugs for IgAN, are approved by the Pharmaceuticals and Medical Devices Agency and MHLW only for the treatment of hypertension or diabetic nephropathy. Therefore, when a patient with IgAN presents with slightly high blood pressure that is not diagnosed as hypertension under regular practice, clinicians may diagnose it as hypertension based on the patient’s medical bill. Moreover, diagnostic codes for IgAN do not necessarily indicate true IgAN because they are not always based on biopsy results. Therefore, the enrollment of patients based solely on the diagnoses recorded on medical bills in clinical research is risky.
In some areas, there have been few reports on the detection of clinical diagnoses from bill data. One such example is type 1 diabetes mellitus (T1DM). Okui et al. reported that patients with T1DM were successfully classified using the bills data [16]. However, most T1DM patients require insulin infusions. In other words, non-use of insulin is a powerful factor in excluding T1DM. Currently, there is no specific treatment for IgAN. This implies that achieving the goal in a manner similar to that in the previous report is impossible.
Several studies have adapted machine learning to big data to identify patients with specific diseases. For example, Nasibeh et al. reported the merits of applying random forest to electronic health record data to detect patients with hypertrophic cardiomyopathy [17]. A random forest was used to produce a predictive model for the clinical diagnosis of hypertrophic cardiomyopathy in each patient using health record data. Machine learning is a form of artificial intelligence that enables computers to recognize inputs and decide what to do to achieve specific goal [18]. Qin et al. reported that patients with CKD with missing clinical data were detected with an accuracy of more than 99% using machine learning technology [19]. Additionally, several studies have developed predictive models for kidney disease using clinical data, which are assumed to be useful for detecting CKD [20–23]. Koyner et al. reported that machine learning had a high performance in predicting acute kidney injury before an increase in serum creatinine levels [24].
However, the aim of these studies was to utilize machine learning as a tool to help clinicians make decisions, and few studies have reported its influence on clinical treatment [25]. To the best of our knowledge, no study has applied machine learning to bill data to detect patients who have already been diagnosed with a specific disease to construct a disease-specific database available for clinical research.
Therefore, we focused on a large public medical insurance system database derived from the nationwide universal healthcare system in Japan. In this study, we aimed to develop a novel method for determining whether an IgAN diagnosis was obtained exclusively from billing data, which were assumed to be utilized for the creation of the disease-specific IgAN cohort from a billing database.
Methods
Data source
This descriptive analytical study enrolled patients from the Department of Nephrology at the University of Tsukuba Hospital. Participants were patients who consulted the department at least once between January 1, 2013, and July 31, 2019. The inclusion and exclusion criteria are shown in Fig 1.
Data extraction
All medical records, including the daily charts and medical bills of the participants, were extracted from the data server of the hospital’s medical ordering system. Billing data included the name of the hospital that generated the bill, the type of insurance the patient used, diagnostic codes for billing, the itemization of the examinations and treatments that the patients received, and an explanation of the reason for any discrepancy between the examinations or treatments and diagnostic codes of each patient.
All collected data were anonymized and virtual IDs were provided to all patients. All data were transformed into numerical forms that indicated the number of times each claim code appeared, except for the binomial form of the presence of the diagnostic codes, during the entire observation period.
Medications were recorded in the form of codes that could only be processed in the electronic bill system, which otherwise had no meaning. These codes were converted into Japanese universal codes provided by the MHLW, known as “MHLW codes” that have regularity in the type of effect of medicines to make it possible to classify patients in clinical characteristics. For example, code 2171005F2021 shows the medicine “COMELIAN KOWA Tablets 100 mg,” a brand of dilazep hydrochloride hydrate. The first 4 digits (“2171”) represent the classification of coronary artery dilators, including the first upper classification of 3 digits (“217,” the vasodilator). The next 3 digits (“005”) represent the specific compound of the classification (dilazep hydrochloride hydrate) with the classification of administration (oral: 000–399; infusion: 400–699; or external: 700–999). The letter next to 7 digits (“F”) represents the form of the medicine (tablet form). The digit next to the letter (“2”) represents the dose of a singular unit (100 mg). The last 3 digits (“021”) represent the unique product of the same form of the same compound (“COMELIAN KOWA”). Using this code, medicines can be classified in specific ways, such as their effects, specific compounds, or specific products.
Review of medical records for the classification of patients
All participants were diagnosed with IgAN during the study period based on a review of their medical records. The diagnosis of IgAN was confirmed when the patient met at least one of the following criteria: 1) description of the presence of diagnosed IgAN in the medical records; 2) diagnosis of IgAN on the pathology report of kidney biopsy; and 3) diagnosis of IgAN on a medical letter from another hospital. Based on this assessment process, specific labels were developed for the presence of IgAN in the patients.
Analysis
Analysis was performed using two methods: manual analysis and machine learning. These two methods were compared to investigate which approach was better for detecting patients with IgAN from medical bill data.
Manual analysis.
Some arbitrary criteria were created by using billing data, based on diagnosis and standard therapy, according to the clinical guidelines and opinions regarding IgAN. The criteria for this process included diagnostic codes, examinations, and treatment of the patients.
The data used for diagnosis and treatment were shaped according to the definitions listed in Table 1. For each of the criteria and their combination, the sensitivity, specificity, positive predictive value (PV+), negative predictive value (PV−), and accuracy for each criterion for the detection of IgAN were analyzed.
In addition, scoring criteria for identifying patients with IgAN in the bill data were manually created. Five factors were selected according to the headings of the components of the Japanese guidelines [26] that are generally considered important in patients with IgAN.
- Two or more serum IgA measurements
- Needle biopsy with immunofluorescence or immunostaining
- Use of angiotensin-converting enzymes inhibitors, direct renin inhibitors, or angiotensin II receptor blockers
- Presence of a group of drugs that includes dilazep hydrochloride hydrate
- Use of any form of corticosteroids
- Treatment of tonsillitis
Based on the number of participants who fulfilled these criteria, ROC analyses were performed.
Machine learning.
The data were analyzed using machine learning. Three datasets were preprocessed using three different methods.
- Raw data: medication and diagnostic codes were not compiled. Bill data were provided directly to the learning program after minimal cleaning, reshaping to a format suitable for the learning program, and removing possible leaky variables.
- Analyzed data with compiled codes of medication and diagnosis, except for diagnostic codes in nephrology: After preprocessing in the same way as in (a), medication and diagnostic codes were compiled in some regularity. Drugs with the same top four digits in the MHLW codes and the same route of administration were placed in the same type of drug. Diagnostic codes recorded as ICD-10 with the same top two letters were compiled into the same diagnosis group. Diagnostic codes in the nephrology category (N0xx–N2xx) were excluded.
- Analyzed data with compiled codes of medication and all diagnoses. In addition to (b), diagnostic codes in nephrology were compiled in the first three letters of the ICD-10.
For the machine-learning processes, the extreme gradient boosting (XGboost) method [27] was selected as the supervised learning method to construct classifiers for the presence of IgAN. There were three reasons for this selection. The first was the size of the dataset, with a maximum number of 3,743 patients, which is considered relatively small for machine learning and does not require excessive machine power. This implies that the fastest processing speed of LightGBM is not required. Second, the data source was a single center with a relatively small sample size. The risk of overfitting is not negligible and should be minimized. The LightGBM creates decision trees with fewer horizontal branches and a deeper hierarchy. According to this characteristic, the risk of overfitting–the excessive adaptation to unimportant details of the learning data–is a concern for LightGBM. Third, from many reports, it has been proved that XGboost practices are useful. Huang et al. reported that, compared to random forest, artificial neural networks, and adaptive boosting, XGboost performed best for predicting heart disease [28]. Moreover, XGBoost has demonstrated high predictive performance for identifying CKD in a clinical database [20]. CatBoost is the newest method that creates many decision trees with numerous patterns of variable weighting [27]; however, there are relatively few reports of this method.
To avoid over- and undertraining, an analysis was conducted using a 5-fold cross-validation method. In each pattern of compilation of diagnostic nephrology codes, as mentioned above, 3,743 patients were divided into five small groups with an equal number (748 or 749) of patients and an equal rate of IgAN patients (nearly 11.7%). Five patterns of allocation of labeled data and data for verification were defined. In each pattern, the representative group was set as the data used for verification, and the other four groups were set as labeled data used for the training process. Thus, each trial of applying XGboost to four groups and verifying it using the remaining group was performed in five patterns. The hyperparameters were set using Optuna–an automated algorism for detecting the optimal tuning of hyperparameters [29]. The representative optimal hyperparameters produced by Optuna are shown in S1 Table.
Evaluation of diagnostic performance.
A receiver operating characteristic (ROC) curve for each manual scoring and machine learning result was constructed, and the area under the curve (AUC) was calculated.
Results
Overview of the data
A total of 3,743 patients met the eligibility criteria shown in Fig 1. An overview of these data is presented in Table 1. As shown in Fig 2, among the 3,743 subjects, 437 (11.7%) were diagnosed with IgAN. Diagnostic codes for IgAN were provided for only 52.9% and 3.2% of patients with and without IgAN, respectively. Sensitivity, specificity, PV+, PV− and accuracy for single and created combined criteria are shown in Table 2. The PV+ value of the IgAN diagnostic codes was 68.5%. Among the patients with IgAN, 13.0% were not assigned any diagnostic codes for glomerular diseases. Moreover, 293 (67.0% of patients with IgAN), 144 (33.0%), and 238 (54.5%) patients received renin-angiotensin system inhibitors, corticosteroids, and a category of vasodilatory agents (including dilazeps), respectively.
Kidney biopsy with immunofluorescence was performed in 36.6% and 11.6% of patients with and without IgAN, respectively.
Analysis of the test performance of manual criteria
As mentioned above, the specificity of the IgAN code was not the highest among the related variables. The three best PV+ are shown in the criteria “any N02x or N03x code combined with tonsil treatment,” “any N02x or N03x code combined with both needle biopsy with immunostaining/immunofluorescence and corticosteroid infusion,” and “code of IgAN.” The PV+ values for each criterion were 1, 0.7031, and 0.6825, respectively. The highest sensitivity (0.6316) was shown in the criterion of “measurement of serum IgA.” However, no single item achieved a coexistence of sensitivity and specificity >80%.
The AUC for manually created criteria was 0.715. Both sensitivity and specificity were below 0.8 at the cutoff set by Youden’s method. The ROC curve is shown in S1 Fig.
Result of machine learning
Fig 3presents the results of plotting the ROC curve of the classifier created by machine learning. AUCs for patterns (a), (b), and (c) were 0.884 ± 0.026, 0.943 ± 0.015, and 0.937 ± 0.014, respectively (Fig 4). The process yielded better results in terms of the AUC in patterns (b) and (c), which used clinically utilized data, than in pattern (a) or manual analysis. The number of variables in pattern (b) (showing the highest average AUC) was 1836. The list of all items is shown in S1 File (in Japanese). The top 10 explanatory variables contributing to efficient classification for each cross-validation group are shown in S2 Table. The explanatory variables that appeared in the top ten in all patterns were age at the time of the oldest data, ICD-10 J3x (acute tonsillitis), ICD-10 N028 (IgAN), and ICD-10 N039 (chronic glomerulonephritis, undifferentiated).
ROC curves using results of machine learning in patterns (a), (b), and (c). (a) The raw data; (b) Analyzed data with compiled codes of medication and diagnosis, except for diagnostic codes in nephrology; (c) Analyzed data with compiled codes of medication and all diagnoses. The group numbers in the legend indicate the group used in verification process (other four groups are used for training in XGboost).
Boxplot of distribution of AUCs for five ROC curves in patterns (a), (b), and (c). (a) The raw data; (b) Data with compiled codes of medication and diagnosis, except for diagnostic codes in nephrology, were analyzed; (c) Data with compiled codes of medication and all diagnoses were analyzed.
Discussion
We tried to identify true patients with IgAN from billing data in view of the inaccuracy of IgAN coding, and subsequently compared the results of manual analysis for the diagnostic performance of IgAN with the results obtained using machine learning. Although a high PV+ was observed in some manual criteria, the sensitivities were extremely low for practical use in all criteria. Considering the low prevalence of IgAN in the general population, a small number of patients were diagnosed using the manual criteria. Thus, machine learning using clinical datasets of medical bills showed far better performance than analysis using the manual method for diagnosing IgAN. We successfully improved the possibility of detecting patients with specific diseases by using medical billing data only. To the best of our knowledge, this is the first report on the use of machine-learning technology to extract true-diagnosed patients from medical billing data without data from biopsy reports, laboratory tests, or clinical records written by physicians.
The strength of these results is their high sensitivity and specificity. The specificities for diagnosing IgAN by machine learning using the utilized datasets surpassed 90% under the set conditions, which is not inferior to the diagnostic code of IgAN alone. Moreover, the sensitivity of machine learning using the utilized datasets was significantly better than that of the IgAN diagnostic code. We believe that this performance would suffice to create an artificial cohort of patients who are strongly predicted to have IgAN, with fewer omissions. It would be possible to establish a virtual cohort of patients that are strongly expected to be diagnosed with IgAN from the medical billing database using this method, which is not susceptible to overdiagnosis or diagnosis of diagnostic codes in the billing data. This technology enables the investigation of the long-term effectiveness of IgAN treatment. Moreover, this methodology can be applied to other diseases in the future.
This study had some limitations. First, the external validity was not guaranteed because the data were collected from a nephrology department at a single center. The direct use of a classifier in other departments or institutions has the potential to cause bias in the local practice of departments, hospitals, or regions. For example, the top 10 factors that were most powerful in distinguishing patients with IgAN from those without IgAN by machine learning, including antiplasmin drugs and carbazochrome sodium sulfonate, were components of the regular clinical pathway applied in our hospital. To solve this problem, billing data from other institutions should be used for the learning or testing processes.
The second limitation is the periodic revision of rules for creating bills. Many codes were added, deleted, or divided into subitems. It is almost impossible to capture all of these changes and reflect them in a single classifier. Thus, some factors used in the analysis may not have been available for different periods. Therefore, it is necessary to cautiously apply the classifier directly to data recorded from different periods. We believe that updating the classifiers using the data extracted during the update period could be a solution to this limitation.
Conclusion
Compared with manual analysis or the application of machine learning to raw billing data, the application of machine-learning technology to well-utilized billing data showed high predictive performance for detecting patients with IgAN. This technology will help construct disease-specific cohorts from the billing data and may contribute to research on various diseases for which it is difficult to conduct a prospective interventional study in nephrology and many other clinical specialties.
Supporting information
S1 Fig. ROC curve plotted by manually constructed scoring based on the related diagnostic codes and treatment for IgAN.
The point on the ROC shows the best cutoff point by Youden’s method (Sensitivity + specificity − 1) and the sensitivity and specificity at the cutoff.
https://doi.org/10.1371/journal.pone.0312915.s001
(TIF)
S1 Table. Representative hyperparameters for XGboost tuned by Optuna.
“eta” defines the learning rate; “round” defines the round number of boosting; “max_depth” defines the maximum depth of the hierarchy of each decision tree; “colsample_bytree” is the probability that each variable is adopted as the source of each decision tree.
https://doi.org/10.1371/journal.pone.0312915.s002
(DOCX)
S2 Table. The top 10 explanatory variables in each group of pattern (b).
F score shows the degrees of contribution of the variable within each group. The absolute values of this score are not comparable straddling different groups. The score does not mean the possibility of IgAN, but the degree of contribution for both inclusion and exclusion of IgAN.
https://doi.org/10.1371/journal.pone.0312915.s003
(DOCX)
S1 File. All components of variables in pattern (b) from extracted bills.
https://doi.org/10.1371/journal.pone.0312915.s004
(DOCX)
References
- 1. Wyatt RJ, Julian BA. IgA nephropathy. N Engl J Med. 2013;368: 2402–2414. pmid:23782179
- 2. McGrogan A, Franssen CFM, de Vries CS. The incidence of primary glomerulonephritis worldwide: a systematic review of the literature. Nephrol Dial Transplant. 2011;26: 414–430. pmid:21068142
- 3. Willey CJ, Coppo R, Schaefer F, Mizerska-Wasiak M, Mathur M, Schultz MJ. The incidence and prevalence of IgA nephropathy in Europe. Nephrol Dial Transplant. 2023;38: 2340–2349. pmid:37156519
- 4. Schena FP, Nistor I. Epidemiology of IgA nephropathy: A global perspective. Semin Nephrol. 2018;38: 435–442. pmid:30177015
- 5. Koyama A, Igarashi M, Kobayashi M. Natural history and risk factors for immunoglobulin A nephropathy in Japan. Research Group on Progressive Renal Diseases. Am J Kidney Dis. 1997;29: 526–532. pmid:9100040
- 6. Kobayashi Y, Tateno S, Hiki Y, Shigematsu H. IgA nephropathy: prognostic significance of proteinuria and histological alterations. Nephron. 1983;34: 146–153. pmid:6348567
- 7. Sato J, Goda K, Kitsuregawa M, Nakashima N, Mitsutake N. Novel analytics framework for universal healthcare insurance database. AMIA Jt Summits Transl Sci Proc. 2019;2019: 353–362. pmid:31258988
- 8. Hirose N, Ishimaru M, Morita K, Yasunaga H. A review of studies using the Japanese National Database of Health Insurance Claims and Specific Health Checkups. Ann Clin Epidemiol. 2020;2: 13–26.
- 9. Fisher ES, Whaley FS, Krushat WM, Malenka DJ, Fleming C, Baron JA, et al. The accuracy of Medicare’s hospital claims data: progress has been made, but problems remain. Am J Public Health. 1992;82: 243–248. pmid:1739155
- 10. Vlasschaert MEO, Bejaimal SAD, Hackam DG, Quinn R, Cuerden MS, Oliver MJ, et al. Validity of administrative database coding for kidney disease: a systematic review. Am J Kidney Dis. 2011;57: 29–43. pmid:21184918
- 11. Lurie N, Popkin M, Dysken M, Moscovice I, Finch M. Accuracy of diagnoses of schizophrenia in Medicaid claims. Hosp Community Psychiatry. 1992;43: 69–71. pmid:1544654
- 12. Bullano MF, Kamat S, Willey VJ, Barlas S, Watson DJ, Brenneman SK. Agreement between administrative claims and the medical record in identifying patients with a diagnosis of hypertension. Med Care. 2006;44: 486–490. pmid:16641668
- 13. Lin C-C, Lai M-S, Syu C-Y, Chang S-C, Tseng F-Y. Accuracy of diabetes diagnosis in health insurance claims data in Taiwan. J Formos Med Assoc. 2005;104: 157–163. pmid:15818428
- 14. Losina E, Barrett J, Baron JA, Katz JN. Accuracy of Medicare claims data for rheumatologic diagnoses in total hip replacement recipients. J Clin Epidemiol. 2003;56: 515–519. pmid:12873645
- 15. Guimarães PO, Krishnamoorthy A, Kaltenbach LA, Anstrom KJ, Effron MB, Mark DB, et al. Accuracy of medical claims for identifying cardiovascular and bleeding events after myocardial infarction: a secondary analysis of the TRANSLATE-ACS study. JAMA Cardiol. 2017;2: 750–757. pmid:28538984
- 16. Ranjbari S, Khatibi T, Vosough Dizaji A, Sajadi H, Totonchi M, Ghaffari F. CNFE-SE: a novel approach combining complex network-based feature engineering and stacked ensemble to predict the success of intrauterine insemination and ranking the features. BMC Med Inform Decis Mak. 2021;21: 1. pmid:33388057
- 17. Farahani NZ, Arunachalam SP, Sundaram DSB, Pasupathy K, Enayati M. Arruda-Olson AM. Explanatory analysis of a machine learning model to identify hypertrophic cardiomyopathy patients from EHR using diagnostic codes. Proceedings (IEEE Int Conf Bioinformatics Biomed). 2020;2020: 1932–1937. pmid:34316386
- 18. Jordan MI, Mitchell TM. Machine learning: trends, perspectives, and prospects. Science. 2015;349: 255–260. pmid:26185243
- 19. Qin J, Chen L, Liu Y, Liu C, Feng C, Chen B. A machine learning methodology for diagnosing chronic kidney disease. IEEE Access. 2019;8: 20991–21002.
- 20. Ogunleye A, Wang Q-G. XGBoost model for chronic kidney disease diagnosis. IEEE ACM Trans Comput Biol Bioinform. 2020;17: 2131–2140. pmid:30998478
- 21. Ma F, Sun T, Liu L, Jing H. Detection and diagnosis of chronic kidney disease using deep learning-based heterogeneous modified artificial neural network. Future Gener Comput Syst. 2020;111: 17–26.
- 22. Anantha Padmanaban KR, Parthiban G. Applying machine learning techniques for predicting the risk of chronic kidney disease. Indian J Sci Technol. 2016;9: 1–6.
- 23. Weber C, Röschke L, Modersohn L, Lohr C, Kolditz T, Hahn U, et al. Optimized identification of advanced chronic kidney disease and absence of kidney disease by combining different electronic health Data Resources and by applying machine learning strategies. J Clin Med. 2020;9. pmid:32932685
- 24. Koyner JL, Carey KA, Edelson DP, Churpek MM. The development of a machine learning inpatient acute kidney injury prediction model. Crit Care Med. 2018;46: 1070–1077. pmid:29596073
- 25. Deo RC. Machine learning in medicine. Circulation. 2015;132: 1920–1930. pmid:26572668
- 26. Yuzawa Y, Yamamoto R, Takahashi K, Katafuchi R, Tomita M, Fujigaki Y, et al. Evidence-based clinical practice guidelines for IgA nephropathy 2014. Clin Exp Nephrol. 2016;20: 511–535. pmid:27095365
- 27. Chen T, Guestrin C. XGboost: A scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining; 2016. pp. 785–794.
- 28. Huang AA, Huang SY. Increasing transparency in machine learning through bootstrap simulation and shapely additive explanations. PLOS ONE. 2023;18: e0281922. pmid:36821544
- 29. Akiba T, Sano S, Yanase T, Ohta T, Koyama M. Optuna: A next-generation hyperparameter optimization framework. In: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining; 2019. pp. 2623–2631.
