Abstract
Objectives
Mass spectrometry-based steroidomics combined with machine learning (ML) provides a potentially powerful approach in endocrine diagnostics, but is hampered by limitations in the conveyance of results and interpretations to clinicians. We address this shortcoming by integration of the two technologies with a laboratory information management systems (LIMS) model.
Methods
The approach involves integration of ML algorithm-derived models with commercially available mathematical programming software and a web-based LIMS prototype. To illustrate clinical utility, the process was applied to plasma steroidomics data from 22 patients tested for primary aldosteronism (PA).
Results
Once mass spectrometry data are uploaded into the system, automated processes enable generation of interpretations of steroid profiles from ML models. Generated reports include plasma concentrations of steroids in relation to age- and sex-specific reference intervals along with results of ML models and narrative interpretations that cover probabilities of PA. If PA is predicted, reports include probabilities of unilateral disease and mutations of KCNJ5 known to be associated with successful outcomes of adrenalectomy. Preliminary results, with no overlap in probabilities of disease among four patients with and 18 without PA and correct classification of all four patients with unilateral PA including three of four with KCNJ5 mutations, illustrate potential utility of the approach to guide diagnosis and subtyping of patients with PA.
Conclusions
The outlined process for integrating plasma steroidomics data and ML with LIMS may facilitate improved diagnostic-decision-making when based on higher-dimensional data otherwise difficult to interpret. The approach is relevant to other diagnostic applications involving ML.
Introduction
Traditional methods for reporting laboratory data for diagnostic decision-making principally involve unidimensional approaches, single measurements of specific analytes reported with reference intervals to establish whether values fall within or outside ranges of populations without the suspected clinical disorder. Even when laboratory tests involve multiple measurements or panels of analytes, the tendency has been to combine measurements into single measures. For example, when screening for primary aldosteronism (PA), the longstanding practice has been to combine measurements of plasma aldosterone and renin into a single ratio [1].
Thinking beyond one dimension in diagnostics is, however, not without associated difficulties. Even when considering test results in the context of a positive vs. negative binary approach, rather than a continuous unidimensional approach, test interpretation can be difficult [2]. Such problems are compounded with multiple measurements.
For endocrine diagnostics, mass spectrometry has become an important laboratory tool for measurements of multiple analytes [3, 4]. In particular, mass spectrometry-based steroid profiling is emerging as a method of considerable potential for diagnostic stratification of patients with adrenal disorders, such as PA, adrenocortical carcinoma and hypercortisolism [5], [6], [7], [8]. However, for any useful interpretation of the multidimensional data provided by such methods, new approaches are required for recognition of patterns in data rather than solely differences of a measurement from predefined reference intervals. Machine learning (ML) is one such approach for interrogating multi-dimensional data and an area of artificial intelligence (AI) that utilizes computational algorithms for different tasks [9, 10]. In diagnostics, these tasks involve classification, which for steroid profiling have been applied broadly to adrenal disorders [8] and more specifically to identification of patients with adrenocortical cancer [11] or PA [12].
Apart from the need to validate ML models, one of several remaining problems for clinical application is how to provide physicians with ML derived data and interpretations in a meaningful and effective manner to facilitate subsequent clinical-decision-making. There is need for processes that allow integration of ML models within the collection of laboratory results and laboratory information management systems (LIMS) for generation of patient laboratory reports. Besides isolated reports on integration of AI with operational systems to improve automation in drug discovery, biomarker identification and other laboratory processes [13, 14], there appears to be little or no coverage in the literature about how AI might be integrated with LIMS to facilitate diagnostic-decision-making.
To address this shortcoming, we outline an approach that we established under a research protocol and clinical trial to prospectively validate previously established ML models for diagnostic stratification of patients with PA. Rather than to further validate ML for this purpose, the objective of the present report is to outline the methods and results of a strategy that we developed to integrate ML with LIMS. To our knowledge, such operational processes remain unreported. Since the clinical protocol is ongoing and involves multiple international centers, it is important to first illustrate the approach using data from patients recruited outside of that clinical trial.
Methods
Overview of implementation
The approach includes integration of ML algorithms with commercially available mathematical programming software and a web-based LIMS model that allows for automated generation of patient reports. The essential elements for the application include an external source, REDCap (Research Electronic Data Capture) tool [15], as a LIMS prototype. These elements interface with MATLAB (MathWorks Inc., Natick, MA) and Microsoft Excel 2019 (Microsoft Inc., Redmond, WA) for processing patient data according to ML algorithm-based models developed and described previously [12] and now undergoing external validation within the PROSALDO (PROspective study on the diagnostic value of Steroid profiling in primary ALDOsteronism) trial.
Patients
To illustrate the application, data for this report were retrieved from 22 patients with hypertension and adrenal adenomas, enrolled under the European Network for the Study of Adrenal Tumors (ENS@T) registry protocol. The ENS@T registry is distinct from the multicenter PROSALDO trial, which was specifically designed for patients with suspected PA. The 22 patients of this report were either recruited before initiation of the PROSALDO trial or failed to meet entry criteria into the trial due to age or other restrictions. The clinical protocol for the ENS@T registry was approved by the Dresden University Hospital Ethics Committee (EK 407122010), and all patients provided written informed consent before their participation.
Data collection and handling
Patient data were collected and managed using REDCap, a secure, web-based software platform designed to support data capture for research studies by way of several features: 1. An intuitive interface for validated data capture; 2. Audit trails for tracking, data manipulation and export procedures; 3. Automated export procedures for seamless data downloads to common software statistical packages; and 4. Procedures for data integration and interoperability with external sources. Through these features, the REDCap platform provides a user-friendly interface for introducing data that can be subsequently integrated and operated with external sources. The automated export functions further allow for download of both data and reports generated from analyses of those data.
By way of the aforementioned functions, the REDCap system hosted at the Technische Universität Dresden was specifically adapted for the PROSALDO trial to enable the upload and processing of mass spectrometry-derived plasma steroid profiles. Following automated data processing for normalization according to variations in age and sex, as well as logarithmic transformations to correct for non-normal distributions, data are uploaded into ML models via external sources. Thereafter, processes are implemented for automated generation of reports that include plasma concentrations of steroids, age- and sex-specific reference intervals and ML results and interpretations.
Steroid profiling
All blood samples for plasma steroid profiling were collected in the morning (08:00–11:00) into blood tubes containing ethylenediaminetetraacetic acid. Separated plasma was stored at −80 °C until analysis. Steroids were determined by liquid chromatography tandem mass spectrometry (LC-MS/MS) using a SCIEX QTRAP 5500 mass spectrometer coupled to a Waters Acquity ultra performance liquid chromatography system (Waters Inc, Milford, MA). The liquid chromatography system included a Phenomenex Kinetex C18 column (2.6 µm, 2.1 × 100 mm) with a column manager, binary solvent manager and a sample manager. Sample preparation first involved protein precipitation, followed by solid phase extraction using OASIS®, HLB-96 well plates and a Positive-Pressure-96 Processor (Waters Inc, Milford, MA) as detailed previously [16]. Validation and assay performance characteristics are also described in that report. The steroid panel included cortisol, 11-deoxycortisol, 21-deoxycortisol, corticosterone, 11-deoxycorticosterone, aldosterone, 18-oxocortisol, 18-hydroxycortisol, cortisone, 17-hydroxyprogesterone, androstenedione, dehydroepiandrosterone (DHEA), DHEA-sulfate (DHEAS) and testosterone. Sex and age-specific reference intervals were derived according to polynomial curve fitting as described previously [17].
Machine learning models
Algorithm-based models were developed as described elsewhere, including details on data preparation, normalization and feature selection [12]. Final models from both learning (training) and internal validation phases were selected based on best performance according to areas under receiver operating characteristic curves and F scores [12]. The algorithms used to generate final ML models included Support Vector Machine (SVM), Random Forest (RF) and Linear Discriminant Analysis (LDA). These three models were selected for further external validation in the PROSALDO trial. All models were built only on steroid concentrations, after normalization with age- and sex-specific reference intervals as described previously [12]. The same normalization is also applied to any new patient data.
Four steroids are included in all models as the main features: aldosterone, 18-oxocortisol 18-hydroxycortisol and 11-deoxycorticosterone. The SVM-based model also includes cortisone, 11-deoxycortisol and androstenedione as additional features, whereas the RF-based model includes 17-hydroxyprogesterone, corticosterone and DHEA as additional features. The LDA model includes 10 instead of seven features and comprises all features of the SVM model plus addition of corticosterone, DHEA and DHEAS.
Sequencing for KCNJ5 mutations
KCNJ5 genotyping was performed by Sanger sequencing of genomic DNA extracted from fresh frozen adrenal nodules, as described previously [18].
Results
Data reception and REDCap interfaces
Data are collected, uploaded, and processed within the REDCap-based prototype for an AI-integrated LIMS with assistance of interfaces to Excel, MATLAB and the mass spectrometry laboratory and clinical study centers from where patient materials and data originate (Figure 1). The workflow begins with manual introduction of clinical data of patients into REDCap by support staff at study centers, which for the purposes of this report was restricted to the Dresden Adrenal Research Center. The same support personnel are responsible for input of data for shipment manifests generated within REDCap after blood samples are collected and processed to plasma specimens for shipment to the mass spectrometry laboratory in Dresden. Manifests also enable samples and results of analyses of those samples to be followed within REDCap by personnel at both study centers and the laboratory.
Flow chart of interfaces and interconnections for integration of ML models with a LIMS model.
Data are collected, uploaded and processed in REDCap with interfaces to Excel and MATLAB. After a patient record is created, shipment manifests are produced and made available electronically to study center staff responsible for transport of samples to the mass spectrometry laboratory. After mass spectrometric analyses, steroid results are uploaded into an Excel template, where ML interpretations run through a MATLAB interface. Results are then uploaded into the REDCap patient record where reports can be reviewed and downloaded.
A REDCap-Excel interface programmed in Visual Basic for Applications (VBA) facilitates download of specific data (dates and times of blood samples, sex, age, patient IDs) from REDCap into an Excel template. Laboratory staff transfer plasma steroid measurements into the template after processing of steroid mass spectral data using MultiQuant™ (AB Sciex, Framingham, MA). The interface uses Windows Hypertext Transfer Protocol (HTTP) Services (WinHTTP) to send POST requests to the Application Programming Interface (API) provided by REDCap. Requests are secured by Transport Layer Security (TLS) and individualized tokens, and are restricted to certain Internet Protocol (IP) addresses or IP address ranges.
The same interface outlined above is also used to upload LC-MS/MS-derived plasma steroid concentrations and ML interpretation into REDCap. WinHTTP POST requests are used to send measurement values to a custom PHP (Hypertext Preprocessor) endpoint running on the Apache web server that hosts REDCap.
MATLAB-facilitated ML interpretations
MATLAB runs on the web server at all times inside a screen environment (Gnu Software, Free Software Foundation, Boston, MA), which facilitates automated application of ML models to generate disease probabilities from steroidomic data. In this way, requests can be processed within seconds as the MATLAB kernel is already running and requires no lengthy startup process with each incoming request. Through the screen environment, MATLAB runs the SVM and RF derived ML models. Results are entered into a text file that is returned by the endpoint via HTTP response to the Excel file (Figure 1).
Data verification and automation process
Steroid profiles and ML results are thereafter uploaded to REDCap, triggering automatic emails to alert review by laboratory personnel qualified to validate the generated data and edit automated narrative reports if required. Automatic email alerts are also triggered if staff intervention is necessary, discrepancies are found, or results are available for download. Reports also contain notifications regarding the review of data. Once the reviewer has approved the data, the reports are automatically generated using the REDCap external module Custom Template Engine. This module inserts uploaded data into a Hypertext Markup Language (HTML) template and generates a downloadable file as a Portable Document Format (PDF) from the HTML template. Immediately after review and verification of reports, investigators at study centers are notified by automated email of the availability of reports for download as PDF files.
Reporting and data visualization
Reports are displayed as visual records with highlighted pathological results outside of reference intervals, ML interpretations and personalized comments (Figure 2). There are three components to the reports. The first component A provides the patient ID and other patient specific data including age, sex as well as sample and report dates and the laboratory contact responsible for validating reports. The second component B, outlines selected steroid concentrations with sex and age matched reference intervals and includes highlighting for values that fall out-side of reference intervals. The third component C includes the ML interpretations, which are displayed in first and last rows according to respective probabilities that the patient does not have vs. does have PA, with an assumption of primary hypertension for the former. The central three rows outline probabilities of bilateral PA, unilateral PA without KCNJ5 mutations and unilateral PA with KCNJ5 mutations. Additional text is provided as automated narrative reports to clarify interpretations according to evaluations by three separate algorithm-based models. Similarly structured reports are also provided for further diagnostic steps corresponding to PA, including the confirmation phase involving saline infusion tests and the subtyping phase involving adrenal venous sampling.
Report sheet for patient number 1.
(A) Provides identification data for the patient (some fields replaced or omitted for confidentiality). (B) Depicts selected steroid concentrations with reference intervals specific to the patient according to sex and age; values out of range of reference intervals are highlighted. (C) Clarifies ML interpretations, including probabilities of primary hypertension vs. primary aldosteronism (PA), displayed as probabilities of bilateral and unilateral PA, with the latter according to presence (Unilateral w/KCNJ5) or absence (Unilateral) of KCNJ5 mutations.
Due to the substantial number of analyzed samples and implicitly, a large number of reports associated with each shipment of patient specimens, the custom template engine module was extended to allow download of ZIP files that contain all PDF reports for a given shipment as identified by the shipment ID. This involves the sending of an API request to the REDCap project that contains the shipment manifests and extracting the pairs of record IDs and study phases for which reports must be created. A ZIP file is created from all generated PDF files using PHP ZipArchive, which can then be downloaded by clinical staff.
Data protection and security
Hypertext Transfer Protocol Secure (HTTPS) is supported and implemented for user login pages. User authentication is handled through Shibboleth (Internet2, Ann Arbor, MI) single sign-on and eduGAIN metadata (eduGAIN, GÉANT Association, Amsterdam, Netherlands). This allows access via existing institutional accounts limited to selected qualified users, with requests to access data secured by TLS. A pseudonymization process with a dual code system, un-related to personal identifiable information, further ensures security of patient data. The system is administered within an internal network, with a secured firewall. Automated database backups are run daily.
Clinical application
Among the 22 patients tested for PA, hyperaldosteronism was excluded in 18 patients according an aldosterone/renin ratio below the cut-off of 31 pmol/mU. Screening in the four other patients revealed a positive aldosterone/renin ratio, a suppressed renin and hypokalemia. Confirmation when necessary involved positive results for a saline infusion test. All four patients had imaging evidence of an adenoma in one adrenal and underwent laparoscopic adrenalectomy of that adrenal. Postoperatively, all four patients had improved blood pressure and remission of hypokalemia. Biochemical cure was confirmed in three patients based on post-operative normalized aldosterone/renin ratios according to the PASO criteria [19], which provided the “gold standard” for confirmation of PA. Post-operative measurements were not possible in the remaining patient due to relocation, but for that patient the adenoma was confirmed to express CYP11B2 by immunohistochemical staining of paraffin embedded adrenal tissue [20].
There was no overlap in ML-based probabilities of PA for patients with and without PA for any of the three algorithms (Table 1). Probabilities of PA ranged from 89 to 100% (median 99%) in patients with PA and from 2 to 90% (median 21%) in those without PA.
Machine learning results depicted as probabilities of primary aldosteronism (PA) in four patients with PA and 18 with PHT.
| No. | Age | Sex | Probabilities of primary aldosteronism | ||
|---|---|---|---|---|---|
| SVM | RF | LDA | |||
| Primary aldosteronism | |||||
| 1 | 44 | F | 92% | 99% | 99% |
| 2 | 39 | M | 89% | 95% | 99% |
| 3 | 60 | F | 100% | 100% | 100% |
| 4 | 51 | F | 95% | 89% | 99% |
|
|
|||||
| Primary hypertension | |||||
|
|
|||||
| 5 | 82 | M | 24% | 10% | 55% |
| 6 | 77 | F | 15% | 11% | 8% |
| 7 | 71 | M | 17% | 20% | 70% |
| 8 | 69 | F | 9% | 2% | 12% |
| 9 | 68 | M | 42% | 70% | 79% |
| 10 | 66 | M | 20% | 30% | 21% |
| 11 | 66 | M | 13% | 8% | 52% |
| 12 | 65 | F | 17% | 16% | 23% |
| 13 | 65 | M | 43% | 49% | 73% |
| 14 | 65 | F | 21% | 12% | 18% |
| 15 | 63 | M | 26% | 77% | 85% |
| 16 | 62 | M | 15% | 11% | 59% |
| 17 | 52 | F | 21% | 1% | 17% |
| 18 | 49 | F | 9% | 36% | 31% |
| 19 | 45 | F | 39% | 32% | 90% |
| 20 | 29 | M | 15% | 8% | 33% |
| 21 | 75 | F | 17% | 23% | 14% |
| 22 | 76 | M | 42% | 62% | 43% |
-
SVM, Support Vector Machine; RF, Random Forest; LDA, Linear Discriminant Analysis.
Among the three different ML algorithm-derived models that are undergoing external validation within the PROSALDO trial, probabilities of PA among patients without PA were -generally lower for the Support Vector Machine (SVM) model than for other models (Table 1). As yet, no cut-offs for these probabilities in terms of predicting PA can be established, though from the available data it is likely that the cut-off for the SVM model will be at or a little above a probability value of 43%, the highest provided among the 20 patients in whom PA was excluded.
ML interpretations indicated unilateral PA in all four patients at probabilities ranging from 73 to 100% (median 91%), including a consistently high likelihood of KCNJ5 mutations in three patients (Table 2). Genotyping of resected nodules revealed mutations of KCNJ5 in those three patients. Available data are too preliminary to establish which ML model is most suitable for diagnostic stratification of patients with PA and what cut-offs might be applied for this purpose.
Machine learning results depicted as probabilities of unilateral primary aldosteronism with (+ve) vs. without (−ve) KCNJ5 mutations vs. bilateral adrenal hyperplasia (BAH) in four patients primary aldosteronism who underwent adrenalectomy.
| SVM | RF | LDA | |||||||
|---|---|---|---|---|---|---|---|---|---|
| No. | +ve | −ve | BAH | +ve | −ve | BAH | +ve | −ve | BAH |
| 1a | 62% | 24% | 8% | 93% | 5% | 1% | 98% | 1% | 0% |
| 2a | 58% | 17% | 14% | 64% | 9% | 21% | 92% | 5% | 1% |
| 3a | 72% | 24% | 4% | 64% | 22% | 14% | 100% | 0% | 0% |
| 4 | 35% | 52% | 8% | 58% | 24% | 7% | 99% | 0% | 0% |
-
aDenotes that a KCNJ5 mutation was detected in the resected adrenal adenoma. See patients 1–4 of Table 1 for the sex and ages. SVM, Support Vector Machine; RF, Random Forest; LDA, Linear Discriminant Analysis.
Discussion
Here we outline an approach on how ML applications of AI may be incorporated into medical LIMS in order to facilitate communication of diagnostic test results and ML interpretations to clinicians. REDCap-based information systems primarily support research rather than the routine laboratory environment. The outlined system was developed for a clinical trial to assess diagnostic utility of ML-based steroidomics models in patients with PA. Other information systems may be more suitable for the routine laboratory environment, but we utilized REDCap because of its user-friendly nature and access to international participants in the diagnostic trial. Custom software integration could also be attained by adapting any middleware software utilized by laboratories rather than an API to realize system connections. If more complex data are required for integration, new software might be developed to extract this information from different sources.
Beyond healthcare, the laboratory informatics, automation, and management software components of LIMS continue to support other applications in fossil fuel, food, beverage, pharmaceutical, engineering and other industries. As in all applications, the history of LIMS in clinical diagnostics can be traced back to the introduction of computerization in the laboratory environment [21, 22]. Modern medical LIMS have evolved from those initially designed to streamline test registration, specimen tracking, data management, reporting and billing to those that now cover more complex and varied tasks to further enhance laboratory efficiency and performance. Such tasks depend on the specific needs of the clinical laboratory and can involve workflows for resource planning, regulatory compliance, biobanking, as well as analyses that involve cytotoxicity assays, proteomics and genomics studies [23], [24], [25], [26], [27].
Although integration of AI within medical LIMS was forecast over 35 years ago [21], this particular task has been slow to crystalize due in part to limitations in the computational power necessary to process the data required to generate and apply ML algorithm-derived models. With the computational power now at hand, there has been an exponential increase in published medical applications of AI. However, there are numerous obstacles to the translation of AI technologies into clinical practice [28], [29], [30], including some relevant to laboratory medicine [29, 31].
Among potential medical applications of AI, those involving biochemical analyses, including mass spectrometry, have shown particularly low translatability [32]. As also reviewed by Herman et al. [10], implementation of AI technologies in the clinical chemistry laboratory, though potentially wide-ranging in scope, has been relatively modest compared to other clinical applications. Nevertheless, despite all obstacles, there has been some progress towards incorporating AI with LIMS to facilitate collection and interpretation of routinely collected data [33], with a clear interest from the commercial sector [34, 35].
Among the various obstacles, the practicalities of integrating ML algorithm-derived models into LIMS are one hurdle to overcome that has received little if any attention in publications related to laboratory medicine. The present report addresses this unmet need. Although specifically directed to integration of LIMS with mass spectrometry-based steroidomics and ML interpretations, similar approaches may be used for other applications of laboratory medicine.
Importantly for any diagnostic application of ML, the algorithm-derived models should have utility for improving not only the diagnostic process, but also therapeutic outcomes [28, 36]. There should be a clear problem to resolve. For PA, the problem is that patients can endure a long, difficult and convoluted multistep diagnostic process that does not always provide a clear path forward to achieve appropriate therapeutic outcomes [37]. As a result of these difficulties, most patients with PA remain undiagnosed [38]. Further compounding the problem, when immunoassays are used, the diagnosis of the bilateral subtype of PA can be more often incorrect than correct [39]. Whether the ML models described here can address these problems requires completion of the PROSALDO trial. In the interim, the preliminary data outlined here, of identification of three of four patients with unilateral PA due to KCNJ5 mutations and complete separation of ML-based probabilities of patients with and without PA, does offer hope for a solution. With such a solution it might be possible to attain a more rapid and accurate diagnosis, which for patients predicted to have somatic KCNJ5 mutations might lead to facilitated surgical intervention without need confirmatory studies and adrenal venous sampling for disease subtyping.
While the purpose of this report is not to validate those models, the presented data not only illustrate how ML models may be integrated with existing and commercially available LIMS, but also the potential utility of combining steroid profiling with ML to assist diagnostic stratification of patients with PA. Beyond this, and as specifically addressed here, we show how ML models can be integrated into a clinical setting, which has been neglected during development of tools for AI [28]. There are, however, other challenges to tackle before the combination of ML with plasma steroidomics can be incorporated into the routine laboratory environment. Foremost is the need to standardize plasma steroid measurements, which is currently being addressed for mass spectrometry-based measurements through international harmonization studies [40]. Data integrity is a pivotal aspect. Transmission of data between laboratory instruments and computer systems should remain accurate, with procedures in place for validation and to maintain data integrity [41]. For mass spectrometry-generated steroidomics data, there remains the need for expert oversight and labor-intensive review of mass spectra. Nevertheless, as described by Herman et al. [10], it is possible that ML may even be applied at this step with auto-verification procedures to streamline transfer of data to LIMS.
Other wider challenges to translating ML into clinical practice relate to confounders, which may propagate an unintentional bias and lead to shortcomings in achieving generalizability to populations beyond those used for training and validation [42]. There are also issues of data security and compliance with emerging regulatory requirements related to applications of AI [43]. The need to generate medical trust in ML models for diagnosis, disease prediction and therapeutic stratification represent other challenges to the technology [36]. However, not all ML models need to be “black boxes” [44, 45]. Transparency about how ML models generate predictions can confer clinical confidence in AI as a tool not to dictate but rather inform clinical decision-making [46]. This can at least in part be attained by appropriate reporting of relevant data with ML interpretations as we outline in this report for interpretations of plasma steroidomics data. Although there are many challenges ahead, none are insurmountable, and as covered in this report that includes the hurdle of integrating AI with LIMS for automated generation of clinical reports containing ML interpretations and other data.
Funding source: Stiftung Charité
Award Identifier / Grant number: BIH_PRO_406
Funding source: Deutsche Forschungsgemeinschaft
Award Identifier / Grant number: 314061271-TRR/CRC 205-1/2
Acknowledgments
Thanks are extended to Denise Kaden and Nicole Hellmig for technical support and to Carola Kunath for support with patients.
-
Research funding: This study was supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation, Projektnummer: 314061271-TRR/CRC 205-1/2 to MP, JWML, TAW and GE and by the Stiftung Charité (BIH_PRO_406 to US).
-
Author contributions: All authors have accepted responsibility for the entire content of this manuscript and approved its submission.
-
Competing interests: MS and GE have a patent pending that has relevance to some aspects of the presented work.
-
Informed consent: Informed consent was obtained from all individuals included in this study.
-
Ethical approval: Research involving human subjects complied with all relevant national regulations, institutional policies and is in accordance with the tenets of the Helsinki Declaration (as revised in 2013), and has been approved by the authors’ Institutional Review Board (EK 407122010).
References
1. Funder, JW, Carey, RM, Mantero, F, Murad, MH, Reincke, M, Shibata, H, et al.. The management of primary aldosteronism: case detection, diagnosis, and treatment: an endocrine society clinical practice guideline. J Clin Endocrinol Metab 2016;101:1889–916. https://doi.org/10.1210/jc.2015-4061.Search in Google Scholar PubMed
2. Morgan, DJ, Pineles, L, Owczarzak, J, Magder, L, Scherer, L, Brown, JP, et al.. Accuracy of practitioner estimates of probability of diagnosis before and after testing. JAMA Intern Med 2021;181:747–55. https://doi.org/10.1001/jamainternmed.2021.0269.Search in Google Scholar PubMed PubMed Central
3. Kushnir, MM, Rockwood, AL, Bergquist, J. Liquid chromatography-tandem mass spectrometry applications in endocrinology. Mass Spectrom Rev 2010;29:480–502. https://doi.org/10.1002/mas.20264.Search in Google Scholar PubMed
4. Ketha, SS, Singh, RJ, Ketha, H. Role of mass spectrometry in clinical endocrinology. Endocrinol Metab Clin N Am 2017;46:593–613. https://doi.org/10.1016/j.ecl.2017.04.001.Search in Google Scholar PubMed
5. Taylor, DR, Ghataore, L, Couchman, L, Vincent, RP, Whitelaw, B, Lewis, D, et al.. A 13-steroid serum panel based on LC-MS/MS: use in detection of adrenocortical carcinoma. Clin Chem 2017;63:1836–46. https://doi.org/10.1373/clinchem.2017.277624.Search in Google Scholar PubMed
6. Wannachalee, T, Turcu, AF. Developments in primary aldosteronism subtyping using steroid profiling. Horm Metab Res 2020;52:373–8. https://doi.org/10.1055/a-1141-3526.Search in Google Scholar PubMed PubMed Central
7. Eisenhofer, G, Masjkur, J, Peitzsch, M, Di Dalmazi, G, Bidlingmaier, M, Grüber, M, et al.. Plasma steroid metabolome profiling for diagnosis and subtyping patients with cushing syndrome. Clin Chem 2018;64:586–96. https://doi.org/10.1373/clinchem.2017.282582.Search in Google Scholar PubMed
8. Wilkes, EH, Rumsby, G, Woodward, GM. Using machine learning to aid the interpretation of urine steroid profiles. Clin Chem 2018;64:1586–95. https://doi.org/10.1373/clinchem.2018.292201.Search in Google Scholar PubMed
9. Richardson, A, Signor, BM, Lidbury, BA, Badrick, T. Clinical chemistry in higher dimensions: machine-learning and enhanced prediction from routine clinical chemistry data. Clin Biochem 2016;49:1213–20. https://doi.org/10.1016/j.clinbiochem.2016.07.013.Search in Google Scholar PubMed
10. Herman, DS, Rhoads, DD, Schulz, WL, Durant, TJS. Artificial intelligence and mapping a new direction in laboratory medicine: a review. Clin Chem 2021;67:1466–82. https://doi.org/10.1093/clinchem/hvab165.Search in Google Scholar PubMed
11. Bancos, I, Taylor, AE, Chortis, V, Sitch, AJ, Jenkinson, C, Davidge-Pitts, CJ, et al.. Urine steroid metabolomics for the differential diagnosis of adrenal incidentalomas in the EURINE-ACT study: a prospective test validation study. Lancet Diabetes Endocrinol 2020;8:773–81. https://doi.org/10.1016/s2213-8587(20)30218-7.Search in Google Scholar
12. Eisenhofer, G, Durán, C, Cannistraci, CV, Peitzsch, M, Williams, TA, Riester, A, et al.. Use of steroid profiling combined with machine learning for identification and subtype classification in primary aldosteronism. JAMA Netw Open 2020;3:e2016209. https://doi.org/10.1001/jamanetworkopen.2020.16209.Search in Google Scholar PubMed PubMed Central
13. Undey, C. AI in process automation. SLAS Technol 2021;26:1–2. https://doi.org/10.1177/2472630320982260.Search in Google Scholar PubMed
14. Marescotti, D, Narayanamoorthy, C, Bonjour, F, Kuwae, K, Graber, L, Calvino-Martin, F, et al.. AI-driven laboratory workflows enable operation in the age of social distancing. SLAS Technol 2022;27:195–203. https://doi.org/10.1016/j.slast.2021.12.001. 35058197.Search in Google Scholar PubMed PubMed Central
15. Harris, PA, Taylor, R, Thielke, R, Payne, J, Gonzalez, N, Conde, JG. Research electronic data capture (REDCap)–a metadata-driven methodology and workflow process for providing translational research informatics support. J Biomed Inf 2009;42:377–81. https://doi.org/10.1016/j.jbi.2008.08.010.Search in Google Scholar PubMed PubMed Central
16. Peitzsch, M, Dekkers, T, Haase, M, Sweep, FC, Quack, I, Antoch, G, et al.. An LC-MS/MS method for steroid profiling during adrenal venous sampling for investigation of primary aldosteronism. J Steroid Biochem Mol Biol 2015;145:75–84. https://doi.org/10.1016/j.jsbmb.2014.10.006.Search in Google Scholar PubMed
17. Eisenhofer, G, Peitzsch, M, Kaden, D, Langton, K, Pamporaki, C, Masjkur, J, et al.. Reference intervals for plasma concentrations of adrenal steroids measured by LC-MS/MS: impact of gender, age, oral contraceptives, body mass index and blood pressure status. Clin Chim Acta 2017;470:115–24. https://doi.org/10.1016/j.cca.2017.05.002.Search in Google Scholar PubMed PubMed Central
18. Scholl, UI, Healy, JM, Thiel, A, Fonseca, AL, Brown, TC, Kunstman, JW, et al.. Novel somatic mutations in primary hyperaldosteronism are related to the clinical, radiological and pathological phenotype. Clin Endocrinol (Oxf) 2015;83:779–89. https://doi.org/10.1111/cen.12873.Search in Google Scholar PubMed PubMed Central
19. Williams, TA, Lenders, JWM, Mulatero, P, Burrello, J, Rottenkolber, M, Adolf, C, et al.. Outcomes after adrenalectomy for unilateral primary aldosteronism: an international consensus on outcome measures and analysis of remission rates in an international cohort. Lancet Diabetes Endocrinol 2017;5:689–99. https://doi.org/10.1016/s2213-8587(17)30135-3.Search in Google Scholar
20. Williams, TA, Gomez-Sanchez, CE, Rainey, WE, Giordano, TJ, Lam, AK, Marker, A, et al.. International histopathology consensus for unilateral primary aldosteronism. J Clin Endocrinol Metab 2021;106:42–54. https://doi.org/10.1210/clinem/dgaa484.Search in Google Scholar PubMed PubMed Central
21. Robinson, AL. LIMS is next step in laboratory automation. Science 1983;220:180–3. https://doi.org/10.1126/science.220.4593.180.Search in Google Scholar PubMed
22. Forest, JC, Rheault, C, Dang-Vu, TK. The laboratory information system (LIS): I-application to the clinical chemistry laboratory. Clin Biochem 1985;18:78–84. https://doi.org/10.1016/s0009-9120(85)80085-0.Search in Google Scholar PubMed
23. Heinen, S, Schulze, N, Franke, B, Klein, F, Lehmann, C, Vehreschild, M, et al.. HEnRY: a DZIF LIMS tool for the collection and documentation of biomaterials in multicentre studies. BMC Bioinf 2020;21:290. https://doi.org/10.1186/s12859-020-03596-1.Search in Google Scholar PubMed PubMed Central
24. Grand, A, Geda, E, Mignone, A, Bertotti, A, Fiori, A. One tool to find them all: a case of data integration and querying in a distributed LIMS platform. Database (Oxford) 2019;2019:1–11. https://doi.org/10.1093/database/baz004.Search in Google Scholar PubMed PubMed Central
25. Faria-Campos, AC, Balottin, LB, Zuin, G, Garcia, V, Batista, PH, Granjeiro, JM, et al.. FluxCTTX: a LIMS-based tool for management and analysis of cytotoxicity assays data. BMC Bioinf 2015;16(19 Suppl):S8. https://doi.org/10.1186/1471-2105-16-s19-s8.Search in Google Scholar
26. Stephan, C, Kohl, M, Turewicz, M, Podwojski, K, Meyer, HE, Eisenacher, M. Using laboratory information management systems as central part of a proteomics data workflow. Proteomics 2010;10:1230–49. https://doi.org/10.1002/pmic.200900420.Search in Google Scholar PubMed
27. Nix, DA, Di Sera, TL, Dalley, BK, Milash, BA, Cundick, RM, Quinn, KS, et al.. Next generation tools for genomic data generation, distribution, and visualization. BMC Bioinf 2010;11:455. https://doi.org/10.1186/1471-2105-11-455.Search in Google Scholar PubMed PubMed Central
28. Lindsell, CJ, Stead, WW, Johnson, KB. Action-informed artificial intelligence-matching the algorithm to the problem. JAMA 2020;323:2141–2. https://doi.org/10.1001/jama.2020.5035.Search in Google Scholar PubMed
29. Matheny, ME, Whicher, D, Thadaney Israni, S. Artificial intelligence in health care: a report from the national academy of medicine. JAMA 2020;323:509–10. https://doi.org/10.1001/jama.2019.21579.Search in Google Scholar PubMed
30. Cohen, IG, Evgeniou, T, Gerke, S, Minssen, T. The European artificial intelligence strategy: implications and challenges for digital health. Lancet Digit Health 2020;2:e376–9. https://doi.org/10.1016/s2589-7500(20)30112-6.Search in Google Scholar PubMed
31. Wieringa, G. Teaching the pony new tricks: competences for specialists in laboratory medicine to meet the challenges of disruptive innovation. Clin Chem Lab Med 2019;57:398–402. https://doi.org/10.1515/cclm-2018-0609.Search in Google Scholar PubMed
32. Eweje, FR, Byun, S, Chandra, R, Hu, F, Kamel, I, Zhang, P, et al.. Translatability analysis of national institutes of health-funded biomedical research that applies artificial intelligence. JAMA Netw Open 2022;5:e2144742. https://doi.org/10.1001/jamanetworkopen.2021.44742.Search in Google Scholar PubMed PubMed Central
33. Liang, H, Tsui, BY, Ni, H, Valentim, CCS, Baxter, SL, Liu, G, et al.. Evaluation and accurate diagnoses of pediatric diseases using artificial intelligence. Nat Med 2019;25:433–8. https://doi.org/10.1038/s41591-018-0335-9.Search in Google Scholar PubMed
34. Labware. Leveraging advanced analytics, metrics and artificial intelligence in your lab. 2021. Available from: https://www.labware.com/lims/ml-ai [Accessed 16 May 2022].Search in Google Scholar
35. Shrestha, A. Machine learning: a primer to laboratory applications; how machine learning could impact the laboratory in the next decade; 2020. Available from: https://www.thermofisher.com/blog/connectedlab/machine-learning-a-primer-to-laboratory-applications.2020 [Accessed 16 May 2022].Search in Google Scholar
36. Gopal, G, Suter-Crazzolara, C, Toldo, L, Eberhardt, W. Digital transformation in healthcare – architectures of present and future information technologies. Clin Chem Lab Med 2019;57:328–35. https://doi.org/10.1515/cclm-2018-0658.Search in Google Scholar PubMed
37. Reincke, M, Bancos, I, Mulatero, P, Scholl, UI, Stowasser, M, Williams, TA. Diagnosis and treatment of primary aldosteronism. Lancet Diabetes Endocrinol 2021;9:876–92. https://doi.org/10.1016/s2213-8587(21)00210-2.Search in Google Scholar
38. Libianto, R, Fuller, PJ, Young, MJ, Yang, J. Primary aldosteronism is a public health issue: challenges and opportunities. J Hum Hypertens 2020;34:478–86. https://doi.org/10.1038/s41371-020-0336-2.Search in Google Scholar PubMed
39. Eisenhofer, G, Kurlbaum, M, Peitzsch, M, Constantinescu, G, Remde, H, Schulze, M, et al.. The saline infusion test for primary aldosteronism: implications of immunoassay inaccuracy. J Clin Endocrinol Metab 2022;107:e2027–36. https://doi.org/10.1210/clinem/dgab924.Search in Google Scholar PubMed PubMed Central
40. Fanelli, F, Cantù, M, Temchenko, A, Mezzullo, M, Lindner, JM, Peitzsch, M, et al.. Report from the HarmoSter study: impact of calibration on comparability of LC-MS/MS measurement of circulating cortisol, 17OH-progesterone and aldosterone. Clin Chem Lab Med 2022;60:726–39. https://doi.org/10.1515/cclm-2021-1028.Search in Google Scholar PubMed
41. Turner, E, Bolton, J. Required steps for the validation of a laboratory information management system. Qual Assur 2001;9:217–24. https://doi.org/10.1080/713844028.Search in Google Scholar PubMed
42. Kelly, CJ, Karthikesalingam, A, Suleyman, M, Corrado, G, King, D. Key challenges for delivering clinical impact with artificial intelligence. BMC Med 2019;17:195. https://doi.org/10.1186/s12916-019-1426-2.Search in Google Scholar PubMed PubMed Central
43. Beckers, R, Kwade, Z, Zanca, F. The EU medical device regulation: implications for artificial intelligence-based medical device software in medical physics. Phys Med 2021;83:1–8. https://doi.org/10.1016/j.ejmp.2021.02.011.Search in Google Scholar PubMed
44. Wongvibulsin, S, Wu, KC, Zeger, SL. Improving clinical translation of machine learning approaches through clinician-tailored visual displays of black box Algorithms: development and validation. JMIR Med Inform 2020;8:e15791. https://doi.org/10.2196/15791.Search in Google Scholar PubMed PubMed Central
45. Cabitza, F, Rasoini, R, Gensini, GF. Unintended consequences of machine learning in medicine. JAMA 2017;318:517–8. https://doi.org/10.1001/jama.2017.7797.Search in Google Scholar PubMed
46. The Lancet Respiratory, M. Opening the black box of machine learning. Lancet Respir Med 2018;6:801. https://doi.org/10.1016/s2213-2600(18)30425-9.Search in Google Scholar
© 2022 Walter de Gruyter GmbH, Berlin/Boston
