gms | German Medical Science

GMDS 2015: 60. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e. V. (GMDS)

Deutsche Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie

06.09. - 09.09.2015, Krefeld

Is clinical chemistry the stepping stone for data mining applications to reach clinical practice?

Meeting Abstract

  • Andreas Bietenbeck - Institut für Klinische Chemie und Pathobiochemie - Klinikum rechts der Isar TU München, München, Deutschland
  • Sabine Gerber - Rechenzentrum - Klinikum rechts der Isar TU München, München, Deutschland
  • Georg Hoffmann - TRILLIUM GmbH, Grafrath, Deutschland
  • Peter Luppa - Institut für Klinische Chemie und Pathobiochemie - Klinikum rechts der Isar TU München, München, Deutschland

GMDS 2015. 60. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e.V. (GMDS). Krefeld, 06.-09.09.2015. Düsseldorf: German Medical Science GMS Publishing House; 2015. DocAbstr. 040

doi: 10.3205/15gmds021, urn:nbn:de:0183-15gmds0212

Published: August 27, 2015

© 2015 Bietenbeck et al.
This is an Open Access article distributed under the terms of the Creative Commons Attribution 4.0 License. See license information at http://creativecommons.org/licenses/by/4.0/.


Outline

Text

Background: In many fields the availability of computational power has changes daily practice profoundly. In banking and insurance complex risk models have largely replaced the experience of the clerks. Most advertising campaigns rely on big data analysis to target potential customers specifically. However, while administration in medicine is largely digitalized, core practice remains unchanged.

Laboratory medicine measures a variety of parameters from patients’ blood and other samples. In Germany analyses are permanently technically and medically validated. For results outside of a predefined reference range warnings can be generated automatically. More complex interpretation of results is left to the laboratory or clinical physician.

The German “Krankenhausentgeltgesetz” (KHEntgG - Hospital Reimbursement Act) requires all hospitals to produce data in a standardized format. This data set includes name, year of birth, gender, insurance, hospital and case ID, day of admission, mode of admission, discharge date, mode of discharge, main and secondary diagnoses coded as ICD-10, operations and medical procedures coded as OPS and other fields.

This work uses these two data sources to estimate the performance of various data mining algorithms. As a generic task the prediction of patients with a long length of stay was chosen.

Methods: The data protection concept was approved by the data-protection supervisor of the Klinkum rechts der Isar. The approval of the local Ethics Committee was granted under the registration number 202/14. Data from the year 2013 was exported from the hospital and laboratory information system. Data from all patients below the age of 18 and from all patients with a mental or behavioral disorder was excluded. For a more secure anonymization, all temporal data fields were shifted. The “Anon-Tool” was used to ensured k-anonymity and l-diversity.

Patients were selected based on the ICD-block that their main diagnosis belongs to. Patients who deceased during their hospital stay or were transferred to another hospital were excluded. To simulate the situation at different points in time, only measurements conducted during the first days in hospital were considered. The allowed number of days was 50%, 70% and 90% of the median length of stay of patients with the same main diagnosis on the three character ICD-code level. Patients who were discharged earlier than 80% of theses cutoffs were excluded.

For preprocessing patients with a low ratio of measured analytes to total analytes and analytes that are only measured in a low ratio of patients were excluded in an iterative process alternatingly. Missing values were imputed to the mean value of the clinical reference range or approximated from other measurements. All values were normalized with the age and gender specific clinical reference range . For each time series of laboratory tests minimum, median, maximum, variance, minimum and maximum slope between each point of measurement and slop of the linear regression through all points were calculated.

Five classification algorithms used these calculated values to predict whether or not a patient belongs to the quintile with the longest length of stay. Quintile calculation was conducted separately for each main diagnosis. The algorithms were One Rule (OneR), Linear Discriminant Analysis with with Recursive Feature Elimination (LDA), Conditional Inference Tree (Ctree), C5.0 and Random forest with Recursive Feature Elimination. All algorithms were evaluated with 5 iterations of 10 fold cross validation.

Results: A total of 5,646,600 measurements from 57,000 hospital visits could be included in the data set after anonymization. The total number of measured parameters from clinical chemistry was 686. On average 28.65 parameters were measured for each patient. A total of 3,847 different main ICD codes were used as main diagnosis. Patients hat 3.99 secondary diagnoses on average. Patients stayed on average 7.3 days in hospital.

C15-C26 “Malignant neoplasms of digestive organs”, I20-I25 “Ischaemic heart diseases” and I60-I69 “Cerebrovascular diseases” were selected as representative ICD-blocks to evaluate the data mining algorithms on. They all contained more than 1,000 patients with a length of stay of 2 days or longer. After data preparation, 12 analytes were included in the analysis for neoplasms of digestive organs, 26 or 27 for ischaemic heart diseases and 20 or 21 for cerebrovascular diseases.

The OneR or Ctree algorithm produced the lowest mean sensitivity and specificity for all but one trial. Random forest and C5.0 provided the best results in the majority of cases. Mean calculated specificities for detection of a long length of stay were always above 0.9 and except for neoplasms of digestive organs the majority of mean calculated specificities were above 0.95. Sensitivities improved from 0.38 to 0.51 (neoplasms of digestive organs), from 0.31 to 0.36 (heart diseases) and from 0.21 to 0.36 (cerebrovascular diseases) as measurement from longer time periods in hospital were included in data analysis. Standard deviation of sensitivity and specificities were always below 0.05, except for one trial with LDA on cerebrovascular diseases data. For all trials, at least one algorithm reached a positive predictive value above 50%. The most important variables were variance of CRP, an inflammation marker, in models for cerebrovascular diseases and a variety of variables loosely connected with anemia for other models.

Discussion: Laboratory medicine measurements and KHEntgG data proved to be a useful source for data analysis. Courses of diseases are highly heterogeneous and only a minority of available analytes is measured for each patient. Generic algorithms classified patients with a long length of stay at different times of hospital stay and for different diseases. Sensitivity of these algorithms is too low to rule out a long length of stay with sufficient certainty. However the positive predictive value seems high enough to justify closer examination of the patients that were predicted to stay longer.

A dialog with treating physicians is needed to identify fields where data mining can influence treatment. Knowledge from these experts can be incorporated into novel predictive models. New features of the patients can be made accessible for data mining to improve performance. Thus, this work can serve as basis for these necessary discussions.