13:00 Uhr
P-03-01:
BioRef: A national infrastructure for generating precise reference intervals for diagnostic medicine
T. Blatter (Bern, CH)
Details anzeigen
Autor:innen:
T. Blatter (Bern, CH)
H. Witte (Bern, CH)
A. Leichtle (Bern, CH)
Introduction/Aim: Reference intervals for patient test results are in standard use across many medical disciplines, allowing physicians to identify potentially pathological test results with relative ease. The process of inferring cohort-specific reference intervals is, however, often ignored due to the high costs and cumbersome effort associated with such a task. Determining reference intervals based on data collected during daily clinical routine using fully automated computational resources may help to lower the associated costs and personalize the reference intervals to the respective cohort population and enhance patient care.
Methods: With the BioRef project, we have developed a multi-center computational framework, where specialized web applications estimate and assess patient group-specific reference intervals based on clinical routine data from four Swiss Hospitals. We have established a common legal governance and interoperability framework for our clinical partners to share their data either to a central database via a national and secure data sharing network or providing their data in a decentralized way via “TI4Health”, a secure and encrypted data-accessing system, allowing each data provider to abide to the restrictions laid out by their ethics waivers.
Results: The deployed web applications, which allow intuitive and interactive data stratification by patient factors (such as age, administrative sex and personal medical history) and laboratory analysis features (such as device, analyzer and test kit identifier) are accessible for registered physicians and researchers. As we are evaluating our deployed framework, we are currently establishing the onboarding of future national and international partners, refining the statistical analysis for multi-cohort patient queries and adjusting the web-interfaces to build clinically viable diagnostic tools.
Conclusion: We present that establishing an opportunity for clinical physicians and researchers to define precise reference intervals in a convenient and reproducible way on-the-fly is a vital part of practicing precision medicine today.
13:06 Uhr
P-03-02:
Analyzing the protein content in human milk for the effects of different pasteurization methods
L. Toll (Villingen-Schwenningen, DE)
Details anzeigen
Autor:innen:
L. Toll (Villingen-Schwenningen, DE)
O. Manzardo (Freiburg, DE)
J. Baumgartner (Freiburg, DE)
E. Nickel (Villingen-Schwenningen, DE)
D. Klotz (Freiburg, DE)
F. Wenzel (Villingen-Schwenningen, DE)
Introduction
In the setup of human milk banking, donated human milk (HM) is frequently pasteurized to reduce potential pathogens and to ensure the safety of premature infants. However, it is known that heat treatment can affect the protein composition of HM. In this work, alterations of selected whey proteins were investigated after the application of three different pasteurization approaches.
Methods
Breast milk samples (n = 15) were Holder-pasteurized (62.5 ± 0.5 °C for 30 min) using either water bath (WB-HoP) or dry temperature (DT-HoP) treatment or were subjected to High Temperature Short Time Treatment (HTST; 62 °C for 5 sec). To enable protein analysis, samples were pretreated by filtration and centrifugation. In the resulting whey, secretory immunoglobulin A (sIgA) and lactoferrin (LF) concentrations were determined by commercially available enzyme-linked immunosorbent assays before and after heat treatment. Alkaline phosphatase activity (ALP) was measured via enzyme activity assay (BioVision, Milpitas, CA, USA).
Results
Both HoP methods resulted in almost a complete decrease of ALP activity (WB-HoP = 0.3 ± 0.4 %, DT-HoP = 0.5 ± 0.4 %), whereas after HTST 52.8 ± 12.2 % was retained (all p < 0.001). The sIgA retention was significantly higher after WB-HoP (73.2 ± 13.5%) and after HTST (80.4 ± 22.7 %) than after DT-HoP (57.0 ± 14.4%, all p < 0.01). In terms of retention of LF, the two HoP methods did not differ significantly (WB-HoP = 47.0 ± 40.0 % vs. DT-HoP = 25.0 ± 9.7 %). Compared to both HoP methods, HTST showed significantly higher retentions of LF (69.9 ± 41.8 %, all p < 0.01).
Conclusion
Holder pasteurization by dry tempering (DT-HoP) seems to have a stronger impact on the quality of human milk than the other two approaches (WB-HoP, HTST). In terms of the protein retention, HTST seems to be a good alternative to the current gold standard HoP.
13:12 Uhr
P-03-03:
Impact of centrifugation and storage conditions on non-fluorescent and fluorescent Nanoparticle Tracking Analysis in plasma blood samples
B. Betz (Jena, DE)
Details anzeigen
Autor:innen:
Z. Ferchichi (Jena, DE)
M. Kiehntopf (Jena, DE)
B. Betz (Jena, DE)
BACKGROUND-AIM
Stability of small extracellular particles (SEP, 45-1000nm) under different pre-analytical conditions is a prerequisite for (Fluorescent) Nanoparticle Tracking Analysis (NTA) in blood samples acquired by health integrated biobanking (HIBB). In this study the impact of different pre-analytical conditions typical for HIBB on (Fluorescence) Nanoparticle Tracking Analysis (NTA) are investigated.
METHODS
Blood was drawn from healthy volunteers in collection tubes with three different anticoagulants (n=10 for each anticoagulant) and centrifuged with 2000g or 4000g. NTA measurement was performed either directly after centrifugation, after storage at room temperature for four hours (2000g) and after freezing at -80C and thawing (2000g). For fluorescent NTA, samples were pre-incubated with the lipophilic fluorescent dye CellMask™ Green.
ZetaView device (Particle Metrix) was used for NTA quantification of small extracellular particles.
RESULTS
Albeit different volunteers donated blood the different anticoagulation tubes, the mean concentration of small extracellular particles was similar in all three tubes after direct measurement indicating a minor influence of the anticoagulant used.
Only in the citrate tube, a (non-significant) drop in SEP concentration after increased centrifugation speed was observed in the non-fluorescent NTA measurement. Although not always significant, there was a clear tendency of loss in SEP concentration after four hours of storage in all the three different collection tube types. There was no loss in SEP concentration after a freeze-thaw cycle.
There was no significant difference in SEP concentration in the fluorescent NTA after increased centrifugation speed, prolonged storage or a freeze-thaw cycle. This was true independent from the tube anticoagulant used.
CONCLUSIONS
The results are to inform researches planning large-scale NTA measurements in samples acquired by HIBB.
In general, NTA in fluorescence mode seems to be less affected by different pre-analytical conditions compared to non-fluorescence NTA and therefore more suitable for HIBB.
13:18 Uhr
P-03-04:
A Two-Step Procedure for Timely and Precise Identification of Hospitalized Diabetic Patients
J. Stolp (Jena, DE)
Details anzeigen
Autor:innen:
J. Stolp (Jena, DE)
M. Kiehntopf (Jena, DE)
B. Betz (Jena, DE)
Background and Introduction
Automated sample collection from diabetic hospitalized patients for biobanking can be challenging due to incomplete patient data at the time of sample selection. This study evaluates a two-step procedure based on machine learning algorithms and natural language processing for timely and precise identification of diabetic patients in the context of healthcare-integrated biobanking.
Method
In the first step of the procedure, logistic regression (LR), and conditional inference forest (CIF) models were trained for the early identification of diabetic individuals on a training dataset (n=550) containing laboratory values from the first 72 hours of hospital stay. Models were then evaluated in a test dataset (n=235) together with a simple rule based laboratory cut-off classifier (LCC). In the second step, laboratory parameters blood glucose and HbA1c, ICD-10 codes or information from discharge summaries extracted by a natural language processing software (NLP-DS) were evaluated for the removal of false positive samples to achieve optimal overall specificity and precission of sample selection.
Results
In the test dataset, evaluation metrics (recall/precision/F1-score) were 71%/86%/0.78 for CIF, 77%/70%/0.74 for LR, and 73%/68%/0.70 for LCC. NLP-DS was the best performing second (review) step (93%/100%/0.97). Combining first-step models with NLP-DS resulted in overall metrics 66%/100%/0.80 for CIF&NLP-DS, 72%/100%/0.84 for LCC&NLP-DS, and 66%/100%/0.80 for LR&NLP-DS. The total case/sample removal rate after the review step was 13% (CIF&NLP-DS), 33% (LR&NLP-DS) and 35% (LCC&NLP-DS).
Conclusion
A machine learning–based two-step procedure seems to be an efficient method for timely identification of diabetic patients and highly precise sample selection enabling targeted automated sample collection for healthcare-integrated biobanking.
13:24 Uhr
P-03-05:
In silico characterization of high-fluorescent cells from cerebrospinal fluid: a supportive diagnostic tool for clinical decision
B. Schwarz (Berlin, DE)
A. Jahic (Berlin, DE)
Details anzeigen
Autor:innen:
B. Schwarz (Berlin, DE)
C. Hardt (Berlin, DE)
M. Prpic (Berlin, DE)
A. Osterloh (Ulm, DE)
F. Heppner (Berlin, DE)
K. Ruprecht (Berlin, DE)
K. Kappert (Berlin, DE)
A. Jahic (Berlin, DE)
Introduction: Automated cytological analysis of cerebrospinal fluid (CSF) represents a convenient platform for accurate and fast quantification of total nucleated cells. These are subsequently classified by a laboratory analyzer as monomorphonuclear (MN: lymphocytes, monocytes), polymorphonuclear (PM: e.g. granulocytes) and high-fluorescent (HF: e.g. plasma cells, erythro-/macrophages, and tumor/malignant cells). CSF HF cells detected on an automated hematology analyzer are currently considered as a research laboratory parameter. Nevertheless, studies have already suggested a diagnostic association with defined clinical patterns. Hence, there is a clinical need for sustained in silico characterization of CSF HF cells combining multiparametric laboratory results with clinical data in well-established patient cohorts.
Aims and Methods: The main goal of the study was to improve the certainty of clinical diagnoses of neurologically associated diseases including HF cells as novel supportive diagnostic criterion. Thereby, we aimed to analyze the diagnostic capability of HF cells characterizing HF cell positive versus HF cell negative cases assigned to various neurologic diagnoses groups such as inflammation, hemorrhage, neoplasia and others.
Results: The results obtained by comparative cytological analysis of manual and automated cell differentiation correlated well for each CSF cell subtype MN, PMN and HF (R2 = 0.85, 0.87 and 0.61, respectively). Diagnosis based correlation values for manual and automated cell differentiation ranged between R2 = 0.66 and 0.95 for MN, R2 = 0.71 and 0.95 for PMN, and R2 = 0.05 and 0.89 for HF showing an insufficient correlation for central nervous system (CNS) associated inflammation (R2 = 0.05), hemorrhage (R2 = 0.22) and other diagnoses (R2 = 0.11) for the HF cell subtype. The correlation coefficient for HF cells in CNS associated neoplasia group was 0.89 indicating diagnostic relevance of HF cells for patients with malignancy. Finally, considering HF cell positive versus HF negative cases underlying a multiparametric analysis with focus on each individual diagnosis group, we assume a supportive predicting and/or diagnostic capability for CSF HF in defined clinical settings.
Conclusion: CSF HF cells detected on a laboratory hematology analyzer seem to have a prognostic and/or diagnostic value for particular diagnoses groups. However, additional laboratory CSF parameters to be included in the clinical decision might further sharpen the multiparametric analysis approach being essential for a reliable diagnostic prediction.
13:30 Uhr
P-03-06:
A general purpose heatmap function in the R environment
J. Spitzer (Bonn, DE)
Details anzeigen
Autor:innen:
J. Spitzer (Bonn, DE)
S. Schmidt (Bonn, DE)
Introduction: Visualisation of high dimensional data remains one of the main challenges in communicating results. Especially in the context of transcriptional data, but also in applications like multiplexed ELISAs, the visual presentation becomes more difficult to manage. For this problem, the heatmap has been established as one of the methods of choice. The ability to display more than 100 variables as well as additional annotational information without creating clutter sets it apart from other methods like bar charts or radar plots. While there are software solutions, they remain poorly optimised and ill fit for customisation. In this project, a software solution has been generated to create heatmaps with a focus on ease of use, customisability, and reproducibility.
Methods: The R programming language is used with a focus on the tidyverse family. The visualisation is created with the ggplot2 package.
Results: Given a table with the data do be visualised, the package will create a heatmap with a default blue-red colour gradient. By default, the first column is expected to be the identifiers for the variables so be plotted on the x-axis, while the rest of non-numeric columns are defaulted to be annotation columns. Rows and columns can be clustered with a variety of distance measures. The returned object can be further customised using the ggplot2 framework; such customisations include different colour gradients, distance measures for clustering, normalisation method and colours for as well as number of annotations.
Conclusion:The software presented here has the goal to drive research through providing an easy to use and highly customisable framework with which to create heatmaps and help in visualisation of high dimensional data. It will support the understandability of high dimensional data and support the interpretation as well as hypothesis generation for any kind of research projects.
13:36 Uhr
P-03-07:
Continuous reference intervals determined with the Shiny application AdRI
S. Klawitter (Wolfenbüttel, DE)
Details anzeigen
Autor:innen:
S. Klawitter (Wolfenbüttel, DE)
G. Hoffmann (Grafrath, DE)
T. Kacprowski (Braunschweig, DE)
F. Klawonn (Braunschweig, DE)
Introduction: Reference intervals are an important part of the interpretation of medical laboratory results. Especially in children and adolescents, their limits sometimes can change very rapidly with age [1]. We suggest continuous methods to better represent the age-dependent progression of reference intervals. A user-friendly Shiny application called AdRI (Age-dependent Reference Intervals), available at https://github.com/SandraKla/AdRI, has been developed for this purpose.
Methods: Generalized additive models for location, scale, and shape parameters (GAMLSS) were developed by Rigby and Stasinopoulos and implemented in the gamlss R package, which provides a variety of features and capabilities for univariate statistical regression modelling and statistical learning [2]. The purpose of the Shiny application AdRI is illustrated using 40 biochemical markers from the Canadian Laboratory Initiative on Pediatric Reference Intervals (CALIPER) [3].
Results: Depending on the additive term used, we obtain different smoothed percentile curves of laboratory values. For alkaline phosphatase (ALP), the GAMLSS with P-splines for females and with Decision Trees for males has the lowest Generalized Akaike Information Criterion (GAIC). All models recognize that the range for ALP of the models is wider after birth and then narrows to follow a relatively constant course over a long period of childhood. With the onset of puberty, the range widens again for both sexes and decreases sharply towards adulthood.
Conclusion: We demonstrate the superiority of continuously modeled reference intervals compared to fixed age groups and provide the Shiny application AdRI to make the technique easily accessible to clinicians and other experts. The influence of pathological values, hyperparameters and the distribution of data over age should be the subject of further investigation, given that the database of laboratory values in newborns and children is generally small.
13:42 Uhr
P-03-08:
A machine learning-derived, blood count based algorithm improves prediction of sepsis
D. Steinbach (Leipzig, DE)
Details anzeigen
Autor:innen:
D. Steinbach (Leipzig, DE)
P. Ahrens (Leipzig, DE)
M. Schmidt (Leipzig, DE)
M. Federbusch (Leipzig, DE)
L. Heuft (Leipzig, DE)
C. Lübbert (Leipzig, DE)
M. Nauck (Greifswald, DE)
M. Gründling (Greifswald, DE)
B. Isermann (Leipzig, DE)
S. Gibb (Grei, DE)
T. Kaiser (Bielefeld, DE)
Introduction: Delay in diagnosing sepsis results in potentially preventable deaths. Mainly due to their complexity or limited applicability, machine learning models to predict sepsis have not yet become part of clinical routines. For this reason, we created a machine learning model that only requires complete blood count (CBC) diagnostics.
Methods: Non-intensive care unit (non-ICU) data from a German tertiary care centre were collected from January 2014 to December 2021. Patient age, sex, and CBC parameters (haemoglobin, platelets, mean corpuscular volume, white and red blood cells) were utilised to train a boosted random forest, which predicts sepsis with ICU admission. Two external validations were conducted using data from the Greifswald University Hospital and the Medical Information Mart for Intensive Care IV database (MIMIC-IV). Using a subset of laboratory data including also procalcitonin, an analogous model was trained with procalcitonin as an additional feature.
Results: After exclusion, 1,381,358 laboratory requests (2016 from sepsis cases) were available. The derived CBC model shows an area under the receiver operating characteristic (AUROC) of 0.872 [CI: 0.857–0.887] for predicting sepsis. External validations show AUROCs of 0.805 [CI: 0.787–0.824] for the Greifswald University Hospital and 0.845 [CI: 0.837–0.852] for MIMIC-IV. The model including procalcitonin revealed a significantly higher performance [AUROC: 0.857; CI: 0.836–0.877] than procalcitonin alone [AUROC: 0.790; CI: 0.759–0.821; p < 0.001]. Thus, the CBC model can facilitate early sepsis prediction in non-ICU patients with high robustness in external validations.
Conclusion: Our results demonstrate that a machine learning-derived algorithm based on routine CBC results improves diagnosis of sepsis, allowing earlier detection of patients at risk. Its implementation in clinical decision support systems has strong potential to provide an essential time advantage and increase patient safety.
13:48 Uhr
P-03-09:
Comparison of electronic acute kidney injury alerts for identification of patients at risk of adverse outcomes
B. Betz (Jena, DE)
Details anzeigen
Autor:innen:
S. Dietsch (Jena, DE)
M. Kiehntopf (Jena, DE)
B. Betz (Jena, DE)
Introduction:
Creatinine based algorithms for electronic alert systems (e-alerts) to detect in-hospital acute kidney injury (AKI) are discussed to be an effective a way for identification of patients at risk of adverse outcomes. A variety of algorithms has been used in studies. Acute kidney disease (AKD) has been described as prolonged kidney damage after acute kidney injury. The association between e-alerts and AKD at hospital discharge has not been studied yet for different algorithms.
Methods:
Creatinine records from a tertiary university hospital were analyzed retrospectively to directly compare five different algorithms for e-alerts. Endpoints were in-hospital mortality or dialysis and AKD at hospital discharge.
Results:
From 6580 hospitalized patients, 2395 (36.4%) had one or more e-alerts. The algorithm that most closely reflects current KDIGO 2012 definition for AKI generated more often (30.9%) and in sum earlier e-alerts as compared to each other algorithm. In addition, the KDIGO-mimicking e-alert had the highest odds ratio for mortality and dialysis during hospital stay before (OR 12.0 CI [9.6-15.0]) and after (OR 6.2 [4.8-8.1]) adjustment for co-variables. After stratification for increase of creatinine (stadium I-III) and combination with another e-alert the modified version of the KDIGO-mimicking e-alert had the highest odds ratio (41.1[25.8-65.7], 57.3[32.8-65.7], 163.0[79.2-335.6] for stadium I,II and III respectively) for AKD at hospital discharge compared to the other e-alerts after adjustment for co-variables.
Conclusion:
All e-alerts can detect patients at increased risk of adverse outcomes during hospital stay (mortality and dialysis) and at discharge from hospital (AKD). A simple modification of the best performing alert (mimicking KDIGO definition for AKI) can further improve risk prediction for AKD at discharge.
13:54 Uhr
P-03-10:
Classifictaion of Mass Spectrometry Profiles of Organic Acids Using Statistical Models
J. Stolp (Jena, DE)
Details anzeigen
Autor:innen:
J. Stolp (Jena, DE)
M. Kiehntopf (Jena , DE)
B. Betz (Jena, DE)
Background and Introduction
The use of statistical models in conjunction with spectral data like infrared spectroscopy or mass spectrometry data has, in medical and scientific contexts, enabled automated classification with high accuracy, precision and recall.
We theorize that machine learning models can be used to analyze mass spectrometry data of organic acids and distinguish between spectra indicating physiological and pathological states.
Methods
79 organic acid mass spectrometry profiles derived from quality control assessment data were used to train and test a random forest model.
Labels for physiological and pathological spectra as well as different disease entities were assigned to the data based on the correct diagnosis according to the respective quality control assessment documents as well as the interpretation of the spectra by a physician.
The data was preprocessed and spectra with large variation of the average area of non disease related sample peaks after normalization to the internal reference standard were excluded.
The data set was balanced by random undersampling and split into training and test data with a ratio of 3 to 1.
A random forest model was designed using the Python Scikit-learn library. Hyperparameters were optimized. The model was trained to classify into physiological and pathological spectra using the training data set and model performance was evaluated using the test data set with accuracy, F1 score, precision and recall as performance metrics and five fold cross validation.
Results
Using quality assesment data an accuracy of 92,8% was reached for the distinction between physiological and pathological spectra using a random forest model. The F1 score reached was 0.92 with a precision of 93% and recall of 87 %.
Conclusion
Machine learning models can be used to analyze mass spectrometry data of organic acids and distinguish between spectra indicating physiological and pathological states with high accuracy and precision.
To improve and validate our model we plan on creating larger datasets using data available from routine measurements of organic acid mass spectrometry profiles.