Predicting Successful Weaning from Mechanical Ventilation by Reduction in Positive End-expiratory Pressure Level Using Machine Learning.

Weaning patients from mechanical ventilation (MV) is a critical and resource intensive process in the Intensive Care Unit (ICU) that impacts patient outcomes and healthcare expenses. Weaning methods vary widely among providers. Prolonged MV is associated with adverse events and higher healthcare expenses. Predicting weaning readiness is a non-trivial process in which the positive end-expiratory pressure (PEEP), a crucial component of MV, has potential to be indicative but has not yet been used as the target. We aimed to predict successful weaning from mechanical ventilation by targeting changes in the PEEP-level using a supervised machine learning model. This retrospective study included 12,153 mechanically ventilated patients from Medical Information Mart for Intensive Care (MIMIC-IV) and eICU collaborative research database (eICU-CRD). Two machine learning models (Extreme Gradient Boosting and Logistic Regression) were developed using a continuous PEEP reduction as target. The data is splitted into 80% as training set and 20% as test set. The model's predictive performance was reported using 95% confidence interval (CI), based on evaluation metrics such as area under the receiver operating characteristic (AUROC), area under the precision-recall curve (AUPRC), F1-Score, Recall, positive predictive value (PPV), and negative predictive value (NPV). The model's descriptive performance was reported as the variable ranking using SHAP (SHapley Additive exPlanations) algorithm. The best model achieved an AUROC of 0.84 (95% CI 0.83-0.85) and an AUPRC of 0.69 (95% CI 0.67-0.70) in predicting successful weaning based on the PEEP reduction. The model demonstrated a Recall of 0.85 (95% CI 0.84-0.86), F1-score of 0.86 (95% CI 0.85-0.87), PPV of 0.87 (95% CI 0.86-0.88), and NPV of 0.64 (95% CI 0.63-0.66). Most of the variables that SHAP algorithm ranked to be important correspond with clinical intuition, such as duration of MV, oxygen saturation (SaO2), PEEP, and Glasgow Coma Score (GCS) components. This study demonstrates the potential application of machine learning in predicting successful weaning from MV based on continuous PEEP reduction. The model's high PPV and moderate NPV suggest that it could be a useful tool to assist clinicians in making decisions regarding ventilator management.

Predicting Hypoxia Using Machine Learning: Systematic Review.

Background: Hypoxia is an important risk factor and indicator for the declining health of inpatients. Predicting future hypoxic events using machine learning is a prospective area of study to facilitate time-critical interventions to counter patient health deterioration. Objective: This systematic review aims to summarize and compare previous efforts to predict hypoxic events in the hospital setting using machine learning with respect to their methodology, predictive performance, and assessed population. Methods: A systematic literature search was performed using Web of Science, Ovid with Embase and MEDLINE, and Google Scholar. Studies that investigated hypoxia or hypoxemia of hospitalized patients using machine learning models were considered. Risk of bias was assessed using the Prediction Model Risk of Bias Assessment Tool. Results: After screening, a total of 12 papers were eligible for analysis, from which 32 models were extracted. The included studies showed a variety of population, methodology, and outcome definition. Comparability was further limited due to unclear or high risk of bias for most studies (10/12, 83%). The overall predictive performance ranged from moderate to high. Based on classification metrics, deep learning models performed similar to or outperformed conventional machine learning models within the same studies. Models using only prior peripheral oxygen saturation as a clinical variable showed better performance than models based on multiple variables, with most of these studies (2/3, 67%) using a long short-term memory algorithm. Conclusions: Machine learning models provide the potential to accurately predict the occurrence of hypoxic events based on retrospective data. The heterogeneity of the studies and limited generalizability of their results highlight the need for further validation studies to assess their predictive performance.

Querying a Clinical Data Warehouse for Combinations of Clinical and Imaging Data.

This study aims to show the feasibility and benefit of single queries in a research data warehouse combining data from a hospital's clinical and imaging systems. We used a comprehensive integration of a production picture archiving and communication system (PACS) with a clinical data warehouse (CDW) for research to create a system that allows data from both domains to be queried jointly with a single query. To achieve this, we mapped the DICOM information model to the extended entity-attribute-value (EAV) data model of a CDW, which allows data linkage and query constraints on multiple levels: the patient, the encounter, a document, and a group level. Accordingly, we have integrated DICOM metadata directly into CDW and linked it to existing clinical data. We included data collected in 2016 and 2017 from the Department of Internal Medicine in this analysis for two query inquiries from researchers targeting research about a disease and in radiology. We obtained quantitative information about the current availability of combinations of clinical and imaging data using a single multilevel query compiled for each query inquiry. We compared these multilevel query results to results that linked data at a single level, resulting in a quantitative representation of results that was up to 112% and 573% higher. An EAV data model can be extended to store data from clinical systems and PACS on multiple levels to enable combined querying with a single query to quickly display actual frequency data.

Preoperative Factors Associated with Target Lesion Revascularization following Endovascular Therapy of the Superficial Femoral Artery.

Objectives: In patients with symptomatic peripheral arterial occlusive disease (PAOD), endovascular revascularization of the superficial femoral artery (SFA) is the most frequent intervention. A major drawback of endovascular procedures is clinically driven target lesion revascularization (CD-TLR), which may cause recurrence of symptoms, re-hospitalizations, and re-interventions. Outcome studies comparing endovascular modalities are heterogeneous and focus more on intraoperative rather than preoperative aspects. Studies have not examined potential risk factors in patients' phenotype before an intervention to prevent CD-TLR. Design: Monocentric, retrospective cohort study of 781 patients with symptomatic PAOD referred to an endovascular intervention of the SFA between 2000 and 2018. Methods: The study aim was to identify risk factors and phenotypes leading to symptomatic PAOD in patients with de novo lesions of the SFA and ≥1 CD-TLR within 12 months post-index procedure. Two groups were differentiated: patients without CD-TLR and with ≥1 CD-TLR. Patient phenotype was compared for cardiovascular (CV) risk factors, age, gender, and renal function. Results: 662 patients (84.8%) (age 73.5 ± 11.2 years; 243 women (36.7%)) with no CD-TLR were compared to 119 patients (15.2%) with ≥1 CD-TLR (age 70.9 ± 12.4 years; 55 women (46.2%)). Women, as well as subjects with dyslipidemia, had each a 1.8-time higher odds ratio of receiving multiple interventions within one year than men or subjects without dyslipidemia. Older subjects (per decade) had a lower odds ratio (0.7) for multiple interventions. Subjects with an eGFR (estimated glomerular filtration rate) <30 mL/min had 3.8 times higher and subjects with eGFR ≥30 and <60 mL/min had a 2.4 higher odds ratio of receiving multiple interventions than subjects with eGFR values ≥90 mL/min. Conclusion: Our data indicate that younger women, patients with dyslipidemia, or those with renal insufficiency are at risk for recurrent midterm CD-TLR after endovascular therapy of the SFA.

Automated provision of clinical routine data for a complex clinical follow-up study: A data warehouse solution.

A deep integration of routine care and research remains challenging in many respects. We aimed to show the feasibility of an automated transformation and transfer process feeding deeply structured data with a high level of granularity collected for a clinical prospective cohort study from our hospital information system to the study's electronic data capture system, while accounting for study-specific data and visits. We developed a system integrating all necessary software and organizational processes then used in the study. The process and key system components are described together with descriptive statistics to show its feasibility in general and to identify individual challenges in particular. Data of 2051 patients enrolled between 2014 and 2020 was transferred. We were able to automate the transfer of approximately 11 million individual data values, representing 95% of all entered study data. These were recorded in n = 314 variables (28% of all variables), with some variables being used multiple times for follow-up visits. Our validation approach allowed for constant good data quality over the course of the study. In conclusion, the automated transfer of multi-dimensional routine medical data from HIS to study databases using specific study data and visit structures is complex, yet viable.

On the way to routine cardiac MRI at 7 Tesla - a pilot study on consecutive 84 examinations.

INTRODUCTION: Cardiac magnetic resonance (CMR) at ultrahigh field (UHF) offers the potential of high resolution and fast image acquisition. Both technical and physiological challenges associated with CMR at 7T require specific hardware and pulse sequences. This study aimed to assess the current status and existing, publicly available technology regarding the potential of a clinical application of 7T CMR. METHODS: Using a 7T MRI scanner and a commercially available radiofrequency coil, a total of 84 CMR examinations on 72 healthy volunteers (32 males, age 19-70 years, weight 50-103 kg) were obtained. Both electrocardiographic and acoustic triggering were employed. The data were analyzed regarding the diagnostic image quality and the influence of patient and hardware dependent factors. 50 complete short axis stacks and 35 four chamber CINE views were used for left ventricular (LV) and right ventricular (RV), mono-planar LV function, and RV fractional area change (FAC). Twenty-seven data sets included aortic flow measurements that were used to calculate stroke volumes. Subjective acceptance was obtained from all volunteers with a standardized questionnaire. RESULTS: Functional analysis showed good functions of LV (mean EF 56%), RV (mean EF 59%) and RV FAC (mean FAC 52%). Flow measurements showed congruent results with both ECG and ACT triggering. No significant influence of experimental parameters on the image quality of the LV was detected. Small fractions of 5.4% of LV and 2.5% of RV segments showed a non-diagnostic image quality. The nominal flip angle significantly influenced the RV image quality. CONCLUSION: The results demonstrate that already now a commercially available 7T MRI system, without major methods developments, allows for a solid morphological and functional analysis similar to the clinically established CMR routine approach. This opens the door towards combing routine CMR in patients with development of advanced 7T technology.

A Specialized System for Arrhythmia Detection for Basic Research in Cardiology.

The detection of cardiac arrhythmias has a long history in medicine, with current developments focusing on early detection using mobile devices. In basic research, however, the use cases and data differ greatly from the experimental setup. We developed a Python-based system to ease detection and analysis of arrhythmic sections in signals measured on extracted and stimulated cardiac myocytes. Multiple algorithms were integrated into the system, tested and evaluated. The best algorithm resulted in an F1-score of 0.97 and was primarily provided in the application.

Improved Lipid Target Level Attainment in Patients with Peripheral Artery Disease.

BACKGROUND: Patients with peripheral artery disease (PAD) fall under the category of a very high cardiovascular risk. Although consequent lipid-lowering therapy (LLT) is advised, only sparse data on attained target level in PAD exists. OBJECTIVES: We aimed to analyse contemporary guideline recommendations for LLT in symptomatic PAD patients. METHODS: A monocentric, prospective, observational study involving 200 symptomatic PAD patients was conducted. Guideline target level attainment and LLT were analysed between 2017 and 2019. RESULTS: Overall, 78.5% of the patients were on statin therapy, mainly of high intensity, with atorvastatin in 50% and rosuvastatin in 33% of the cases. The average statin dosage adjusted for simvastatin was 55 mg/d. Low density lipoprotein-cholesterol (LDL-C) was <1.8 mmol/L in 53% and <1.4 mmol/L in 34% of the cases. Mean LDL-C levels were at 1.85 ± 0.88 mmol/L. We observed no difference in the treatment and the target level attainment of patients with a stable PAD (intermittent claudication) or chronic critical PAD. However, patients with ≥ 1 vascular region affected (i.e., coronary and/or cerebrovascular) were treated more intensively and had lower LDL-C levels than patients with PAD alone. CONCLUSION: It appears that there are more awareness and improvement of previously documented undertreatment of LDL-C levels in symptomatic PAD patients. Although statin treatment is initiated in the majority of patients, our findings call for a continuously intensified LLT in symptomatic PAD patients.

Adherence to statin therapy favours survival of patients with symptomatic peripheral artery disease.

AIMS: We hypothesized that adherence to statin therapy determines survival in patients with peripheral artery disease (PAD). METHODS AND RESULTS: Single-centre longitudinal observational study with 691 symptomatic PAD patients. Mortality was evaluated over a mean follow-up of 50 ± 26 months. We related statin adherence and low-density lipoprotein cholesterol (LDL-C) target attainment to all-cause mortality. Initially, 73% of our PAD patients were on statins. At follow-up, we observed an increase to 81% (P < 0.0001). Statin dosage, normalized to simvastatin 40 mg, increased from 50 to 58 mg/day (P < 0.0001), and was paralleled by a mean decrease of LDL-C from 97 to 82 mg/dL (P < 0.0001). The proportion of patients receiving a high-intensity statin increased over time from 38% to 62% (P < 0.0001). Patients never receiving statins had a significant higher mortality rate (31%) than patients continuously on statins (13%) or having newly received a statin (8%; P < 0.0001). Moreover, patients on intensified statin medication had a low mortality of 9%. Those who terminated statin medication or reduced statin dosage had a higher mortality (34% and 20%, respectively; P < 0.0001). Multivariate analysis showed that adherence to or an increase of the statin dosage (both P = 0.001), as well as a newly prescribed statin therapy (P = 0.004) independently predicted reduced mortality. CONCLUSION: Our data suggest that adherence to statin therapy is associated with reduced mortality in symptomatic PAD patients. A strategy of intensive and sustained statin therapy is recommended.

Unlocking the PACS DICOM Domain for its Use in Clinical Research Data Warehouses.

Clinical Data Warehouses (DWHs) are used to provide researchers with simplified access to pseudonymized and homogenized clinical routine data from multiple primary systems. Experience with the integration of imaging and metadata from picture archiving and communication systems (PACS), however, is rare. Our goal was therefore to analyze the viability of integrating a production PACS with a research DWH to enable DWH queries combining clinical and medical imaging metadata and to enable the DWH to display and download images ad hoc. We developed an application interface that enables to query the production PACS of a large hospital from a clinical research DWH containing pseudonymized data. We evaluated the performance of bulk extracting metadata from the PACS to the DWH and the performance of retrieving images ad hoc from the PACS for display and download within the DWH. We integrated the system into the query interface of our DWH and used it successfully in four use cases. The bulk extraction of imaging metadata required a median (quartiles) time of 0.09 (0.03-2.25) to 12.52 (4.11-37.30) seconds for a median (quartiles) number of 10 (3-29) to 103 (8-693) images per patient, depending on the extraction approach. The ad hoc image retrieval from the PACS required a median (quartiles) of 2.57 (2.57-2.79) seconds per image for the download, but 5.55 (4.91-6.06) seconds to display the first and 40.77 (38.60-41.63) seconds to display all images using the pure web-based viewer. A full integration of a production PACS with a research DWH is viable and enables various use cases in research. While the extraction of basic metadata from all images can be done with reasonable effort, the extraction of all metadata seems to be more appropriate for subgroups.

Development of a 3-Dimensional Prognostic Score for Patients With Symptomatic Peripheral Artery Disease: PAD3D Score.

Peripheral artery disease (PAD) is a high-risk condition for cardiovascular (CV) events, but no specific prognosis assessment tool exists. We developed an individual risk score (PAD3D) based on the combined predictive value for mortality, including (1) age, (2) severity of PAD, and (3) extent of atherosclerosis. Patients (n = 1310) with symptomatic PAD were followed up for a mean of 50 ± 26 months. The cohort was randomly subdivided into a test and validation cohort. All-cause and CV mortality were prospectively analyzed for PAD3D score and in combination with classical risk factors. For the test and validation cohort (n = 655 each), all-cause and CV mortality were predicted (P < .001) by the PAD3D score. Additional inclusion of classical risk factors did not increase discrimination compared with PAD3D as "area under receiver-operating characteristic" curves were similar for both scores at any time point. Thus, the addition of the classical risk factors to PAD3D did not further improve the prognostic value. The PAD3D score provides a risk gradient of a 4.5-fold increase in all-cause and CV mortality. We developed a score for precise prediction of all-cause and CV mortality. The PAD3D score promises to allow for personalized goals in risk intervention.

Implementation of a HL7-CQL Engine Using the Graph Database Neo4J.

The Clinical Quality Language (CQL) is a useful tool for defining search requests for data stores containing FHIR data. Unfortunately, there are only few execution engines that are able to evaluate CQL queries. As FHIR data represents a graph structure, the authors pursue the approach of storing all data contained in a FHIR server in the graph database Neo4J and to translate CQL queries into Neo4J's query language Cypher. The query results returned by the graph database are retranslated into their FHIR representation and returned to the querying user. The approach has been positively tested on publicly available FHIR servers with a handcrafted set of example CQL queries.

Mitofusin 2 Is Essential for IP3-Mediated SR/Mitochondria Metabolic Feedback in Ventricular Myocytes.

Aim: Endothelin-1 (ET-1) and angiotensin II (Ang II) are multifunctional peptide hormones that regulate the function of the cardiovascular and renal systems. Both hormones increase the intracellular production of inositol-1,4,5-trisphosphate (IP3) by activating their membrane-bound receptors. We have previously demonstrated that IP3-mediated sarcoplasmic reticulum (SR) Ca2+ release results in mitochondrial Ca2+ uptake and activation of ATP production. In this study, we tested the hypothesis that intact SR/mitochondria microdomains are required for metabolic IP3-mediated SR/mitochondrial feedback in ventricular myocytes. Methods: As a model for disrupted mitochondrial/SR microdomains, cardio-specific tamoxifen-inducible mitofusin 2 (Mfn2) knock out (KO) mice were used. Mitochondrial Ca2+ uptake, membrane potential, redox state, and ATP generation were monitored in freshly isolated ventricular myocytes from Mfn2 KO mice and their control wild-type (WT) littermates. Results: Stimulation of ET-1 receptors in healthy control myocytes increases mitochondrial Ca2+ uptake, maintains mitochondrial membrane potential and redox balance leading to the enhanced ATP generation. Mitochondrial Ca2+ uptake upon ET-1 stimulation was significantly higher in interfibrillar (IFM) and perinuclear (PNM) mitochondria compared to subsarcolemmal mitochondria (SSM) in WT myocytes. Mfn2 KO completely abolished mitochondrial Ca2+ uptake in IFM and PNM mitochondria but not in SSM. However, mitochondrial Ca2+ uptake induced by beta-adrenergic receptors activation with isoproterenol (ISO) was highest in SSM, intermediate in IFM, and smallest in PNM regions. Furthermore, Mfn2 KO did not affect ISO-induced mitochondrial Ca2+ uptake in SSM and IFM mitochondria; however, enhanced mitochondrial Ca2+ uptake in PNM. In contrast to ET-1, ISO induced a decrease in ATP levels in WT myocytes. Mfn2 KO abolished ATP generation upon ET-1 stimulation but increased ATP levels upon ISO application with highest levels observed in PNM regions. Conclusion: When the physical link between SR and mitochondria by Mfn2 was disrupted, the SR/mitochondrial metabolic feedback mechanism was impaired resulting in the inability of the IP3-mediated SR Ca2+ release to induce ATP production in ventricular myocytes from Mfn2 KO mice. Furthermore, we revealed the difference in Mfn2-mediated SR-mitochondrial communication depending on mitochondrial location and type of communication (IP3R-mRyR1 vs. ryanodine receptor type 2-mitochondrial calcium uniporter).

Query Translation Between AQL and CQL.

Secondary use of electronic health records using data aggregation systems (DAS) with standardized access interfaces (e.g. openEHR, i2b2, FHIR) have become an attractive approach to support clinical research. In order to increase the volume of underlying patient data, multiple DASs at different institutions can be connected to research networks. Two obstacles to connect a DAS to such a network are the syntactical differences between the involved DAS query interfaces and differences in the data models the DASs operate on. The current work presents an approach to tackle both problems by translating queries from a DAS using openEHR's query language AQL (Archetype Query Language) into queries using the query language CQL (Clinical Quality Language) and vice versa. For the subset of queries which are expressible in both query languages the presented approach is well feasible.

Query Translation Between openEHR and i2b2.

Secondary use of electronic health records using data warehouses (DW) has become an attractive approach to support clinical research. In order to increase the volume of underlying patient data DWs at different institutions can be connected to research networks. Two obstacles to connect a DW to such a network are the syntactical differences between the involved DW technologies and differences in the data models of the connected DWs. The current work presents an approach to tackle both problems by translating queries from the DW system openEHR into queries from the DW system i2b2 and vice versa. For the subset of queries expressible in the query languages of both systems, the presented approach is well feasible.

Replicating medication trend studies using ad hoc information extraction in a clinical data warehouse.

BACKGROUND: Medication trend studies show the changes of medication over the years and may be replicated using a clinical Data Warehouse (CDW). Even nowadays, a lot of the patient information, like medication data, in the EHR is stored in the format of free text. As the conventional approach of information extraction (IE) demands a high developmental effort, we used ad hoc IE instead. This technique queries information and extracts it on the fly from texts contained in the CDW. METHODS: We present a generalizable approach of ad hoc IE for pharmacotherapy (medications and their daily dosage) presented in hospital discharge letters. We added import and query features to the CDW system, like error tolerant queries to deal with misspellings and proximity search for the extraction of the daily dosage. During the data integration process in the CDW, negated, historical and non-patient context data are filtered. For the replication studies, we used a drug list grouped by ATC (Anatomical Therapeutic Chemical Classification System) codes as input for queries to the CDW. RESULTS: We achieve an F1 score of 0.983 (precision 0.997, recall 0.970) for extracting medication from discharge letters and an F1 score of 0.974 (precision 0.977, recall 0.972) for extracting the dosage. We replicated three published medical trend studies for hypertension, atrial fibrillation and chronic kidney disease. Overall, 93% of the main findings could be replicated, 68% of sub-findings, and 75% of all findings. One study could be completely replicated with all main and sub-findings. CONCLUSION: A novel approach for ad hoc IE is presented. It is very suitable for basic medical texts like discharge letters and finding reports. Ad hoc IE is by definition more limited than conventional IE and does not claim to replace it, but it substantially exceeds the search capabilities of many CDWs and it is convenient to conduct replication studies fast and with high quality.

Information Extraction from Echocardiography Reports for a Clinical Follow-up Study-Comparison of Extracted Variables Intended for General Use in a Data Warehouse with Those Intended Specifically for the Study.

BACKGROUND: The interest in information extraction from clinical reports for secondary data use is increasing. But experience with the productive use of information extraction processes over time is scarce. A clinical data warehouse has been in use at our university hospital for several years, which also provides an information extraction of echocardiography reports developed for general use. OBJECTIVES: This study aims to illustrate the difficulties encountered, while using data from a preexisting information extraction process for a large clinical study. To compare the data from the preexisting process with the data obtained from a specially developed process designed to improve the quality and completeness of the study data. METHODS: We extracted the echocardiography variables for 440 patients from the general-use information extraction of the data warehouse (678 reports). Then we developed an information extraction process for the same variables but specifically for this study, with the aim to extract as much information as possible from the text. The extracted data of both processes were compared with a newly created gold standard defined by a cardiologist with long-standing experience in heart failure. RESULTS: Among 57 echocardiography variables considered relevant for the study, 50 were documented in the routine text reports and could be extracted. Twenty of the required variables were not provided by the general-use extraction process, some others were not provided correctly. The median macro F1-score (precision, recall) across the 30 variables for which values were extracted was 0.81 (0.94, 0.77). Across all 50 variables, as relevant for the study, median macro F1-score was only 0.49 (0.56, 0.46). Employing the study-specific approach considerably improved the quality and completeness of the variables, resulting in F1-scores of 0.97 (0.98, 0.96) across all variables. CONCLUSION: Data from information extractions can be used for large clinical studies. However, preexisting information extraction processes should be treated with caution, as the time and effort spent defining each variable in the information extraction process may not be clear.

Non-invasive ultrasound-based imaging of atherosclerosis.

Early detection of vascular damage in atherosclerosis and accurate assessment of cardiovascular risk factors are the basis for appropriate treatment strategies in cardiovascular medicine. The current review focuses on non-invasive ultrasound-based methods for imaging of atherosclerosis. Endothelial dysfunction is an accepted early manifestation of atherosclerosis. The most widely used technique to study endothelial function is non-invasive, flow-mediated dilation of the brachial artery under high-resolution ultrasound imaging. Although an increased intima-media thickness value is associated with future cardiovascular events in several large population studies, systematic use is not recommended in clinical practice for risk assessment of individual persons. Carotid plaque analysis with grey-scale median, 3-D ultrasound or contrast-enhanced ultrasound are promising techniques for further scientific work in prevention and therapy of generalized atherosclerosis.

Finding Needles in the Haystack: Identifying Patients with Rare Subtype of Multiple Myeloma Supported by a Data Warehouse and Information Extraction.

Finding patient cases with extremely rare pathologies is a laborious task. To decrease time spent on manually searching through thousands of discharge letters and reports, a data warehouse with a fast fulltext search index was queried. Our use case is to find "macrofocal myeloma", i.e. Multiple Myeloma patients with few large lesions. We guessed the number of those patients in the University Hospital Würzburg at about 20. Most criteria were available in the data warehouse in an unstructured form requiring information extraction. 8 patient cases were found by searching for different spellings of "macrofocal myeloma" in discharge letters directly. With an indirect search combining several criteria, we found additional 23 candidate patient cases, from which 10 were classified by a domain expert as correct. The most difficult criteria were determining the degree of bone marrow infiltration. We achieved an F1 score of 93.2 % for this task. The number of patient cases to be screened manually for this disease decreased from about 25000 to 23.