Band 48, Heft 5 - Special Issue: At the Crossroads between Statistics and Artificial Intelligence: Statistical Learning in Laboratory Medicine; Issue Editors: Georg Hoffmann and Frank Klawonn

Objectives The multifaceted potential of ChatGPT in the medical domain remains underexplored, particularly regarding its application in software development by individuals with a medical background but limited information technology expertise. Methods This study investigates ChatGPT’s utility in creating a laboratory medicine application. Results Despite minimal programming skills, the authors successfully developed an automated intra-assay, inter-device precision test for immunophenotyping with a shiny user interface, facilitated by ChatGPT. While the coding process was expedited, meticulous oversight and error correction by the authors were imperative. Conclusions These findings highlight the value of large language models such as ChatGPT in code-based application development for automating work processes in a medical context. Particularly noteworthy is the facilitation of these tasks for non-technically trained medical professionals and its potential for digital medical education.

Objectives The study aims to acquaint readers with six widely used machine learning (ML) techniques (Principal Component Analysis (PCA), Uniform Manifold Approximation and Projection (UMAP), k-means, hierarchical clustering and the decision tree models (rpart and random forest)) that might be useful for the analysis of laboratory data. Methods Utilizing a recently validated data set from lung cancer diagnostics, we investigate how ML can support the search for a suitable tumor marker panel for the differentiation of small cell (SCLC) and non-small cell lung cancer (NSCLC). Results The ML techniques used here effectively helped to gain a quick overview of the data structures and provide initial answers to the clinical questions. Dimensionality reduction techniques such as PCA and UMAP offered insightful visualization and impression of the data structure, suggesting the existence of two tumor groups with a large overlap of largely inconspicuous values. This impression was confirmed by a cluster analysis with the k-means algorithm, indicative of unsupervised learning. For supervised learning, decision tree models like rpart or random forest demonstrated their utility in differential diagnosis of the two tumor types. The rpart model, which constructs binary decision trees based on the recursive partitioning algorithm, suggests a tree involving four serum tumor markers (STMs), which were confirmed by the random forest approach. Both highlighted pro-gastrin-releasing peptide (ProGRP), neuron specific enolase (NSE), cytokeratin-19 fragment (CYFRA 21-1) and cancer antigen (CA) 72-4 as key tumor markers, aligning with the outcomes of the initial statistical analysis. Cross-validation of the two proposals showed a higher area under the receiver operating characteristic (AUROC) curve of 0.95 with a 95 % confidence interval (CI) of 0.92–0.97 for the random forest model compared to an AUROC curve of 0.88 (95 % CI: 0.83–0.93). Conclusions ML can provide a useful overview of inherent medical data structures and distinguish significant from less pertinent features. While by no means replacing human medical and statistical expertise, ML can significantly accelerate the evaluation of medical data, supporting a more informed diagnostic dialogue between physicians and statisticians.

Objectives In the context of exploratory data analysis and machine learning, standardization of laboratory results is an important pre-processing step. Variable proportions of pathological results in routine datasets lead to changes of the mean ( µ ) and standard deviation ( σ ), and thus cause problems in the classical z-score transformation. Therefore, this study investigates whether the zlog transformation compensates these disadvantages and makes the results more meaningful from a medical perspective. Methods The results presented here were obtained with the statistical software environment R, and the underlying data set was obtained from the UC Irvine Machine Learning Repository. We compare the differences of the zlog and z-score transformation for five different dimension reduction methods, hierarchical clustering and four supervised classification methods. Results With the zlog transformation, we obtain better results in this study than with the z-score transformation for dimension reduction, clustering and classification methods. By compensating the disadvantages of the z-score transformation, the zlog transformation allows more meaningful medical conclusions. Conclusions We recommend using the zlog transformation of laboratory results for pre-processing when exploratory data analysis and machine learning techniques are applied.

Objectives Reference intervals (RI) play a decisive role in the interpretation of medical laboratory results. An important step in the determination of RI is age- and sex specific partitioning, which is usually based on an empirical approach by graphical representation. In this study, we evaluate an automated machine learning approach. Methods This study uses pediatric data from the CALIPER RI (Canadian Laboratory Initiative on Pediatric Reference Intervals) study. The calculation of potential partitions is carried out using a regression tree model included in the rpart package of the statistical programming language R. The Harris & Boyd method is used to compare the corresponding partitions suggested by rpart and CALIPER. For better comparability, the reference ranges of the partitions of both approaches are then calculated using reflimR. Results Most of the partitions suggested by rpart or CALIPER show sufficient heterogeneity among themselves to justify age- and/or sex-specific RI partitioning. With only few individual exceptions, both methods yield comparable results. The partitions of both approaches for albumin and γ-glutamyltransferase are very similar to each other. For creatinine rpart suggests a slightly earlier distinction between the sexes. Alkaline phosphatase shows the most pronounced differences. In addition to a considerable earlier sex split, rpart suggests different age intervals for both sexes, resulting in three partitions for females and four partitions for males. Conclusions Our findings indicate that the automated analysis provided by rpart yields results that comparable to traditional methods. Nevertheless, the medical plausibility of the automatic suggestions needs to be validated by human experts.

Objectives Reference intervals for the general clinical practice are expected to cover non-pathological values, but also reflect the underlying biological variation present in age- and gender-specific patient populations. Reference intervals can be inferred from routine patient data measured in high capacity using parametric approaches. Stratified reference distributions are obtained which may be transformed to normality via e.g. a Yeo-Johnson transformation. The estimation of the optimal transformation parameter for Yeo-Johnson through maximum likelihood can be highly influenced by the presence of outlying observations, resulting in biased reference interval estimates. Methods To reduce the influence of outlying observations on parametric reference interval estimation, a reweighted M-estimator approach for the Yeo-Johnson (YJ) transformation was utilised to achieve central normality in stratified reference populations for a variety of laboratory test results. The reweighted M-estimator for the YJ transformation offers a robust parametric approach to infer relevant reference intervals. Results The proposed method showcases robustness up to 15 % of outliers present in routine patient data, highlighting the applicability of the reweighted M-estimator in laboratory medicine. Furthermore, reference intervals are personalised based on the patients’ age and gender for a variety of analytes from routine patient data collected in a tertiary hospital, robustly reducing the dimensionality of the data for more data-driven approaches. Conclusions The method shows the advantages for estimating reference intervals directly and parametrically from routine patient data in order to provide expected reference ranges. This approach to locally inferred reference intervals allows a more nuanced comparison of patients’ test results.

Objectives Although there are several indirect methods that can be used to verify reference limits, they have a common weakness in that they assume a low proportion of pathological values. This paper investigates whether a Gaussian decomposition algorithm can identify the non-pathological fraction even if it is not the main subset of mixed data. Methods All investigations are carried out in the R programming environment. The mclust package is used for Gaussian mixture modelling via the expectation maximization (EM) algorithm. For right-skewed distributions, logarithms of the original values are taken to approximate the Gaussian model. We use the Bayesian information criterion (BIC) for evaluation of the results. The reflimR and refineR packages serve as comparison procedures. Results We generate synthetic data mixtures with known normal distributions to demonstrate the feasibility and reliability of our approach. Application of the algorithm to real data from a Nigerian and a German population produces results, which help to interpret reference intervals of reflimR and refineR that are obviously too wide. In the first example, the mclust analysis of hemoglobin in Nigerian women supports the medical hypothesis that an anemia rate of more than 50 % leads to falsely low reference limits. Our algorithm proposes various scenarios based on the BIC values, one of which suggests reference limits that are close to published data for Nigeria but significantly lower than those established for the Caucasian population. In the second example, the standard statistical analysis of creatine kinase in German patients with predominantly cardiac diseases yields a reference interval that is clearly too wide. With mclust we identify overlapping fractions that explain this false result. Conclusions Gaussian mixture modelling does not replace standard methods for reference interval estimation but is a valuable adjunct when these methods produce discrepant or implausible results.