Abstract: Machine learning (ML) is rapidly emerging in several fields of cancer research. ML algorithms can deal with vast amounts of medical data and provide a better understanding of malignant disease. Its ability to process information from different diagnostic modalities and functions to predict prognosis and suggest therapeutic strategies indicates that ML is a promising tool for the future management of hematologic malignancies; acute myeloid leukemia (AML) is a model disease of various recent studies. An integration of these ML techniques into various applications in AML management can assure fast and accurate diagnosis as well as precise risk stratification and optimal therapy. Nevertheless, these techniques come with various pitfalls and need a strict regulatory framework to ensure safe use of ML. This comprehensive review highlights and discusses recent advances in ML techniques in the management of AML as a model disease of hematologic neoplasms, enabling researchers and clinicians alike to critically evaluate this upcoming, potentially practice-changing technology.

Abstract: The tyrosine-protein phosphatase nonreceptor type 11 (PTPN11) is an important regulator of RAS signaling and frequently affected by mutations in patients with acute myeloid leukemia (AML). Despite the relevance for leukemogenesis and as a potential therapeutic target, the prognostic role is controversial. To investigate the prognostic impact of PTPN11 mutations, we analyzed 1529 adult AML patients using next-generation sequencing. PTPN11 mutations were detected in 106 of 1529 (6.93%) patients (median VAF: 24%) in dominant (36%) and subclonal (64%) configuration. Patients with PTPN11 mutations were associated with concomitant mutations in NPM1 (63%), DNMT3A (37%), and NRAS (21%) and had a higher rate of European LeukemiaNet (ELN) favorable cytogenetics (57.8% vs 39.1%; P & lt; .001) and higher white blood cell counts (P = .007) compared with PTPN11 wild-type patients. In a multivariable analysis, PTPN11 mutations were independently associated with poor overall survival (hazard ratio [HR]: 1.75; P & lt; .001), relapse-free survival (HR: 1.52; P = .013), and a lower rate of complete remission (odds ratio: 0.46; P = .008). Importantly, the deleterious effect of PTPN11 mutations was confined predominantly to the ELN favorable-risk group and patients with subclonal PTPN11 mutations (HR: 2.28; P & lt; .001) but not found with dominant PTPN11 mutations (HR: 1.07; P = .775), presumably because of significant differences within the rate and spectrum of associated comutations. In conclusion, our data suggest an overall poor prognostic impact of PTPN11 mutations in AML, which is significantly modified by the underlying cytogenetics and the clonal context in which they occur.

Abstract: The evaluation of bone marrow morphology by experienced hematopathologists is essential in the diagnosis of acute myeloid leukemia (AML); however, it suffers from a lack of standardization and inter-observer variability. Deep learning (DL) can process medical image data and provides data-driven class predictions. Here, we apply a multi-step DL approach to automatically segment cells from bone marrow images, distinguish between AML samples and healthy controls with an area under the receiver operating characteristic (AUROC) of 0.9699, and predict the mutation status of Nucleophosmin 1 ( NPM1 )—one of the most common mutations in AML—with an AUROC of 0.92 using only image data from bone marrow smears. Utilizing occlusion sensitivity maps, we observed so far unreported morphologic cell features such as a pattern of condensed chromatin and perinuclear lightening zones in myeloblasts of NPM1 -mutated AML and prominent nucleoli in wild-type NPM1 AML enabling the DL model to provide accurate class predictions.

Abstract: Extramedullary manifestations (EM) are rare in acute myeloid leukemia (AML) and their impact on clinical outcomes is controversially discussed. Methods We retrospectively analyzed a large multi-center cohort of 1583 newly diagnosed AML patients, of whom 225 (14.21%) had EM. Results AML patients with EM presented with significantly higher counts of white blood cells ( p   〈  0.0001), peripheral blood blasts ( p   〈  0.0001), bone marrow blasts ( p  = 0.019), and LDH ( p   〈  0.0001). Regarding molecular genetics, EM AML was associated with mutations of NPM1 (OR: 1.66, p   〈  0.001), FLT3 -ITD (OR: 1.72, p   〈  0.001) and PTPN11 (OR: 2.46, p   〈  0.001). With regard to clinical outcomes, EM AML patients were less likely to achieve complete remissions (OR: 0.62, p  = 0.004), and had a higher early death rate (OR: 2.23, p  = 0.003). Multivariable analysis revealed EM as an independent risk factor for reduced overall survival (hazard ratio [HR]: 1.43, p   〈  0.001), however, for patients who received allogeneic hematopoietic cell transplantation (HCT) survival did not differ. For patients bearing EM AML, multivariable analysis unveiled mutated TP53 and IKZF1 as independent risk factors for reduced event-free (HR: 4.45, p   〈  0.001, and HR: 2.05, p  = 0.044, respectively) and overall survival (HR: 2.48, p  = 0.026, and HR: 2.63, p  = 0.008, respectively). Conclusion Our analysis represents one of the largest cohorts of EM AML and establishes key molecular markers linked to EM, providing new evidence that EM is associated with adverse risk in AML and may warrant allogeneic HCT in eligible patients with EM.

Abstract: Acute promyelocytic leukemia (APL) is considered a hematologic emergency due to high risk of bleeding and fatal hemorrhages being a major cause of death. Despite lower death rates reported from clinical trials, patient registry data suggest an early death rate of 20%, especially for elderly and frail patients. Therefore, reliable diagnosis is required as treatment with differentiation-inducing agents leads to cure in the majority of patients. However, diagnosis commonly relies on cytomorphology and genetic confirmation of the pathognomonic t(15;17). Yet, the latter is more time consuming and in some regions unavailable. Methods In recent years, deep learning (DL) has been evaluated for medical image recognition showing outstanding capabilities in analyzing large amounts of image data and provides reliable classification results. We developed a multi-stage DL platform that automatically reads images of bone marrow smears, accurately segments cells, and subsequently predicts APL using image data only. We retrospectively identified 51 APL patients from previous multicenter trials and compared them to 1048 non-APL acute myeloid leukemia (AML) patients and 236 healthy bone marrow donor samples, respectively. Results Our DL platform segments bone marrow cells with a mean average precision and a mean average recall of both 0.97. Further, it achieves high accuracy in detecting APL by distinguishing between APL and non-APL AML as well as APL and healthy donors with an area under the receiver operating characteristic of 0.8575 and 0.9585, respectively, using visual image data only. Conclusions Our study underlines not only the feasibility of DL to detect distinct morphologies that accompany a cytogenetic aberration like t(15;17) in APL, but also shows the capability of DL to abstract information from a small medical data set, i. e. 51 APL patients, and infer correct predictions. This demonstrates the suitability of DL to assist in the diagnosis of rare cancer entities. As our DL platform predicts APL from bone marrow smear images alone, this may be used to diagnose APL in regions were molecular or cytogenetic subtyping is not routinely available and raise attention to suspected cases of APL for expert evaluation.

Abstract: Achievement of complete remission (CR) signifies a crucial milestone in the therapy of acute myeloid leukemia (AML) while refractory disease is associated with dismal outcomes. Hence, accurately identifying patients at risk is essential to tailor treatment concepts individually to disease biology. Machine Learning (ML) is a branch of computer science that can process large data sets for a plethora of purposes. The underlying mechanism does not necessarily begin with a manually drafted hypothesis model. Rather the ML algorithms can detect patterns in pre-processed data and derive abstract information. We used ML to predict CR and 2-year overall survival (OS) in a large multi-center cohort of 1383 AML patients who received intensive induction therapy using clinical, laboratory, cytogenetic and molecular genetic data. To enable a customizable and reusable technological approach and achieve optimal results, we designed a data-driven platform with an embedded, automated ML pipeline integrating state-of-the-art software technology for data management and ML models. The platform consists of five scalable modules for data import and modelling, data transformation, model refinement, machine learning algorithms, feature support and performance feedback that are executed in an iterative manner to approach step-wisely the optimal configuration. To reduce dimensionality and the the risk of overfitting, dynamic feature selection was used, i.e. features were selected according to their support by feature selection algorithms. To be included in an ML model, a feature had to pass a pre-determined threshold of overall predictive power determined by summing the normalized scores of the feature selection algorithms. Features below the threshold were automatically excluded from the ML models for the respective iteration. In that way, features of high redundancy or low entropy were automatically filtered out. Our classification algorithms were completely agnostic of pre-existing risk classifications and autonomously selected predictive features both including established markers of favorable or adverse risk as well as identifying markers of so-far controversial relevance. De novo AML, extramedullary AML, double-mutated (dm) CEBPA, mutations of CEBPA-bZIP, NPM1, FLT3-ITD, ASXL1, RUNX1, SF3B1, IKZF1, TP53, U2AF1, t(8;21), inv(16)/t(16;16), del5/del5q, del17, normal or complex karyotypes, age and hemoglobin at initial diagnosis were statistically significant markers predictive of CR while t(8;21), del5/del5q, inv(16)/t(16;16), del17, dm CEBPA, CEBPA-bZIP, NPM1, FLT3-ITD , DNMT3A, SF3B1, U2AF1, TP53, age, white blood cell count, peripheral blast count, serum LDH and Hb at initial diagnosis as well as extramedullary manifestations were predictive for 2-year OS. For prediction of CR and 2-year OS, AUROCs ranged between 0.77 - 0.86 and 0.63 - 0.74, respectively. We provide a method to automatically select predictive features from different data types, cope with gaps and redundancies, apply and optimize different ML models, and evaluate optimal configurations in a scalable and reusable ML platform. In a proof-of-concept manner, our algorithms utilize both established markers of favorable or adverse risk and also provide further evidence for the roles of U2AF1, IKZF1, SF3B1, DNMT3A and bZIP mutations of CEBPA in AML risk prediction. Our study serves as a fundament for prospective validation and data-driven ML-guided risk assessment in AML at initial diagnosis for the individual patient. Image caption: Patient features were automatically selected by machine learning to predict complete remission (CR) and 2-year overall survival (OS) after intensive induction therapy. Based on a continuous feature support metric with a predefined cut-off of 0.5 (determined by optimal classification performance), 27 and 25 features were automatically selected for prediction of CR (A) and 2-year OS (C), respectively. For each of these features predicted by machine learning, odds ratios and 95% confidence intervals (CI) were calculated for CR (B) and 2 year OS (D). BMB: bone marrow blast count; FLT3h/low: FLT3-ITD ratio, h=high & gt;0.5; Hb: hemoglobin; karyotype, c: complex aberrant karyotype (≥ 3 aberrations); karyotype, n: normal karyotype (no aberrations); LDH: lactate dehydrogenase; PBB: peripheral blood blast count; PLT: platelet count; WBC: white blood cell count. Figure 1 Figure 1. Disclosures Schetelig: Roche: Honoraria, Other: lecture fees; Novartis: Honoraria, Other: lecture fees; BMS: Honoraria, Other: lecture fees; Abbvie: Honoraria, Other: lecture fees; AstraZeneca: Honoraria, Other: lecture fees; Gilead: Honoraria, Other: lecture fees; Janssen: Honoraria, Other: lecture fees . Platzbecker: Janssen: Honoraria; Celgene/BMS: Honoraria; AbbVie: Honoraria; Novartis: Honoraria; Takeda: Honoraria; Geron: Honoraria. Müller-Tidow: Pfizer: Research Funding; Janssen: Consultancy, Research Funding; Bioline: Research Funding. Baldus: Celgene/BMS: Honoraria; Amgen: Honoraria; Novartis: Honoraria; Jazz: Honoraria. Krause: Siemens: Research Funding; Takeda: Honoraria; Pfizer: Honoraria; art-tempi: Honoraria; Kosmas: Honoraria; Gilead: Other: travel support; Abbvie: Other: travel support. Haenel: Bayer Vital: Honoraria; Jazz: Consultancy, Honoraria; GSK: Consultancy; Takeda: Consultancy, Honoraria; Novartis: Consultancy, Honoraria; Roche: Consultancy, Honoraria; Amgen: Consultancy; Celgene: Consultancy, Honoraria. Schliemann: Philogen S.p.A.: Consultancy, Honoraria, Research Funding; Abbvie: Consultancy, Other: travel grants; Astellas: Consultancy; AstraZeneca: Consultancy; Boehringer-Ingelheim: Research Funding; BMS: Consultancy, Other: travel grants; Jazz Pharmaceuticals: Consultancy, Research Funding; Novartis: Consultancy; Roche: Consultancy; Pfizer: Consultancy. Middeke: Roche: Consultancy, Honoraria; Janssen: Consultancy, Honoraria, Research Funding; Abbvie: Consultancy, Honoraria; Pfizer: Consultancy, Honoraria; Jazz: Consultancy; Astellas: Consultancy, Honoraria; Sanofi: Honoraria, Research Funding; Novartis: Consultancy; Gilead: Consultancy; Glycostem: Consultancy; UCB: Honoraria.

Abstract: In cancer diagnostics, a considerable amount of data is acquired during routine work-up. Recently, machine learning has been used to build classifiers that are tasked with cancer detection and aid in clinical decision-making. Most of these classifiers are based on supervised learning (SL) that needs time- and cost-intensive manual labeling of samples by medical experts for model training. Semi-supervised learning (SSL), however, works with only a fraction of labeled data by including unlabeled samples for information abstraction and thus can utilize the vast discrepancy between available labeled data and overall available data in cancer diagnostics. In this review, we provide a comprehensive overview of essential functionalities and assumptions of SSL and survey key studies with regard to cancer care differentiating between image-based and non-image-based applications. We highlight current state-of-the-art models in histopathology, radiology and radiotherapy, as well as genomics. Further, we discuss potential pitfalls in SSL study design such as discrepancies in data distributions and comparison to baseline SL models, and point out future directions for SSL in oncology. We believe well-designed SSL models to strongly contribute to computer-guided diagnostics in malignant disease by overcoming current hinderances in the form of sparse labeled and abundant unlabeled data.

Abstract: Precision oncology is grounded in the increasing understanding of genetic and molecular mechanisms that underly malignant disease and offer different treatment pathways for the individual patient. The growing complexity of medical data has led to the implementation of machine learning techniques that are vastly applied for risk assessment and outcome prediction using either supervised or unsupervised learning. Still largely overlooked is reinforcement learning (RL) that addresses sequential tasks by exploring the underlying dynamics of an environment and shaping it by taking actions in order to maximize cumulative rewards over time, thereby achieving optimal long-term outcomes. Recent breakthroughs in RL demonstrated remarkable results in gameplay and autonomous driving, often achieving human-like or even superhuman performance. While this type of machine learning holds the potential to become a helpful decision support tool, it comes with a set of distinctive challenges that need to be addressed to ensure applicability, validity and safety. In this review, we highlight recent advances of RL focusing on studies in oncology and point out current challenges and pitfalls that need to be accounted for in future studies in order to successfully develop RL-based decision support systems for precision oncology.

Abstract: With an urgent need for bedside imaging of coronavirus disease 2019 (COVID-19), this study’s main goal was to assess inter- and intraobserver agreement in lung ultrasound (LUS) of COVID-19 patients. In this single-center study we prospectively acquired and evaluated 100 recorded ten-second cine-loops in confirmed COVID-19 intensive care unit (ICU) patients. All loops were rated by ten observers with different subspeciality backgrounds for four times by each observer (400 loops overall) in a random sequence using a web-based rating tool. We analyzed inter- and intraobserver variability for specific pathologies and a semiquantitative LUS score. Interobserver agreement for both, identification of specific pathologies and assignment of LUS scores was fair to moderate (e.g., LUS score 1 Fleiss’ κ = 0.27; subpleural consolidations Fleiss’ κ = 0.59). Intraobserver agreement was mostly moderate to substantial with generally higher agreement for more distinct findings (e.g., lowest LUS score 0 vs. highest LUS score 3 (median Fleiss’ κ = 0.71 vs. 0.79) or air bronchograms (median Fleiss’ κ = 0.72)). Intraobserver consistency was relatively low for intermediate LUS scores (e.g. LUS Score 1 median Fleiss’ κ = 0.52). We therefore conclude that more distinct LUS findings (e.g., air bronchograms , subpleural consolidations ) may be more suitable for disease monitoring, especially with more than one investigator and that training material used for LUS in point-of-care ultrasound (POCUS) should pay refined attention to areas such as B-line quantification and differentiation of intermediate LUS scores.