Ground truth labels challenge the validity of sepsis consensus definitions in critical illness.

BACKGROUND: Sepsis is the leading cause of death in the intensive care unit (ICU). Expediting its diagnosis, largely determined by clinical assessment, improves survival. Predictive and explanatory modelling of sepsis in the critically ill commonly bases both outcome definition and predictions on clinical criteria for consensus definitions of sepsis, leading to circularity. As a remedy, we collected ground truth labels for sepsis. METHODS: In the Ground Truth for Sepsis Questionnaire (GTSQ), senior attending physicians in the ICU documented daily their opinion on each patient's condition regarding sepsis as a five-category working diagnosis and nine related items. Working diagnosis groups were described and compared and their SOFA-scores analyzed with a generalized linear mixed model. Agreement and discriminatory performance measures for clinical criteria of sepsis and GTSQ labels as reference class were derived. RESULTS: We analyzed 7291 questionnaires and 761 complete encounters from the first survey year. Editing rates for all items were > 90%, and responses were consistent with current understanding of critical illness pathophysiology, including sepsis pathogenesis. Interrater agreement for presence and absence of sepsis was almost perfect but only slight for suspected infection. ICU mortality was 19.5% in encounters with SIRS as the "worst" working diagnosis compared to 5.9% with sepsis and 5.9% with severe sepsis without differences in admission and maximum SOFA. Compared to sepsis, proportions of GTSQs with SIRS plus acute organ dysfunction were equal and macrocirculatory abnormalities higher (p < 0.0001). SIRS proportionally ranked above sepsis in daily assessment of illness severity (p < 0.0001). Separate analyses of neurosurgical referrals revealed similar differences. Discriminatory performance of Sepsis-1/2 and Sepsis-3 compared to GTSQ labels was similar with sensitivities around 70% and specificities 92%. Essentially no difference between the prevalence of SIRS and SOFA ≥ 2 yielded sensitivities and specificities for detecting sepsis onset close to 55% and 83%, respectively. CONCLUSIONS: GTSQ labels are a valid measure of sepsis in the ICU. They reveal suspicion of infection as an unclear clinical concept and refute an illness severity hierarchy in the SIRS-sepsis-severe sepsis spectrum. Ground truth challenges the accuracy of Sepsis-1/2 and Sepsis-3 in detecting sepsis onset. It is an indispensable intermediate step towards advancing diagnosis and therapy in the ICU and, potentially, other health care settings.

Leveraging implicit expert knowledge for non-circular machine learning in sepsis prediction.

Sepsis is the leading cause of death in non-coronary intensive care units. Moreover, a delay of antibiotic treatment of patients with severe sepsis by only few hours is associated with increased mortality. This insight makes accurate models for early prediction of sepsis a key task in machine learning for healthcare. Previous approaches have achieved high AUROC by learning from electronic health records where sepsis labels were defined automatically following established clinical criteria. We argue that the practice of incorporating the clinical criteria that are used to automatically define ground truth sepsis labels as features of severity scoring models is inherently circular and compromises the validity of the proposed approaches. We propose to create an independent ground truth for sepsis research by exploiting implicit knowledge of clinical practitioners via an electronic questionnaire which records attending physicians' daily judgements of patients' sepsis status. We show that despite its small size, our dataset allows to achieve state-of-the-art AUROC scores. An inspection of learned weights for standardized features of the linear model lets us infer potentially surprising feature contributions and allows to interpret seemingly counterintuitive findings.

De-Identification of German Medical Admission Notes.

Medical texts are a vast resource for medical and computational research. In contrast to newswire or wikipedia texts medical texts need to be de-identified before making them accessible to a wider NLP research community. We created a prototype for German medical text de-identification and named entity recognition using a three-step approach. First, we used well known rule-based models based on regular expressions and gazetteers, second we used a spelling variant detector based on Levenshtein distance, exploiting the fact that the medical texts contain semi-structured headers including sensible personal data, and third we trained a named entity recognition model on out of domain data to add statistical capabilities to our prototype. Using a baseline based on regular expressions and gazetteers we could improve F2-score from 78% to 85% for de-identification. Our prototype is a first step for further research on German medical text de-identification and could show that using spelling variant detection and out of domain trained statistical models can improve de-identification performance significantly.

Automated Classification of Selected Data Elements from Free-text Diagnostic Reports for Clinical Research.

OBJECTIVES: In the Multiple Myeloma clinical registry at Heidelberg University Hospital, most data are extracted from discharge letters. Our aim was to analyze if it is possible to make the manual documentation process more efficient by using methods of natural language processing for multiclass classification of free-text diagnostic reports to automatically document the diagnosis and state of disease of myeloma patients. The first objective was to create a corpus consisting of free-text diagnosis paragraphs of patients with multiple myeloma from German diagnostic reports, and its manual annotation of relevant data elements by documentation specialists. The second objective was to construct and evaluate a framework using different NLP methods to enable automatic multiclass classification of relevant data elements from free-text diagnostic reports. METHODS: The main diagnoses paragraph was extracted from the clinical report of one third randomly selected patients of the multiple myeloma research database from Heidelberg University Hospital (in total 737 selected patients). An EDC system was setup and two data entry specialists performed independently a manual documentation of at least nine specific data elements for multiple myeloma characterization. Both data entries were compared and assessed by a third specialist and an annotated text corpus was created. A framework was constructed, consisting of a self-developed package to split multiple diagnosis sequences into several subsequences, four different preprocessing steps to normalize the input data and two classifiers: a maximum entropy classifier (MEC) and a support vector machine (SVM). In total 15 different pipelines were examined and assessed by a ten-fold cross-validation, reiterated 100 times. For quality indication the average error rate and the average F1-score were conducted. For significance testing the approximate randomization test was used. RESULTS: The created annotated corpus consists of 737 different diagnoses paragraphs with a total number of 865 coded diagnosis. The dataset is publicly available in the supplementary online files for training and testing of further NLP methods. Both classifiers showed low average error rates (MEC: 1.05; SVM: 0.84) and high F1-scores (MEC: 0.89; SVM: 0.92). However the results varied widely depending on the classified data element. Preprocessing methods increased this effect and had significant impact on the classification, both positive and negative. The automatic diagnosis splitter increased the average error rate significantly, even if the F1-score decreased only slightly. CONCLUSIONS: The low average error rates and high average F1-scores of each pipeline demonstrate the suitability of the investigated NPL methods. However, it was also shown that there is no best practice for an automatic classification of data elements from free-text diagnostic reports.