Acute Myeloid Leukemia (AML) is a highly aggressive hematological malignancy characterized by the infiltration of cancer cells in the bone marrow. It is associated with lower remission rates as patients age, and the average overall survival rate ranges from 12 to 18 months (Rose-Inman and Kuehl, 2014; Pelcovits and Niroula, 2020).

The European Leukemia Net (ELN) established guidelines for diagnosing and treating AML in 2010 (Döhner et al., 2010). These served as a cornerstone in the field, providing valuable insights. Subsequent updates were published in 2017 (Döhner et al., 2017) and 2022 (Döhner et al., 2022), reflecting advancements in understanding AML's biomarkers, disease subtypes, and overall behavior. These updates have contributed to a more comprehensive and up-to-date disease management approach.

According to the current diagnostic criteria for AML, established by the World Health Organization (WHO), the presence of at least 10 or 20% myeloblasts in the bone marrow or peripheral blood is required, depending on the specific molecular subtype of the disease (Arber et al., 2022). These guidelines, outlined in the Classification of Tumours of Haematopoietic and Lymphoid Tissues, provide standardized criteria for diagnosing AML accurately.

Apart from the initial diagnosis, patients with AML also undergo a prognostic evaluation to determine their risk profile, typically categorized into favorable, intermediate, and adverse. This risk stratification relies on analyzing cytogenetic and molecular characteristics (The Cancer Genome Atlas Research Network, 2013). Cytogenetic characteristics involve specific chromosome alterations, while mutations in genes such as NPM1, RUNX1, ASXL1, TP53, BCOR, EZH2, SF3B1, SRSF2, STAG2, and ZRSR2 determine molecular characteristics. Healthcare professionals extensively employ the ELN risk classification to make critical treatment decisions, as it directly influences the patient's prognosis, quality of life, and overall survival.

The main problem with the current ELN risk classification is the significant variability within the same risk group. Accurately assessing the intermediate-risk group is especially challenging, potentially causing delays in starting treatment and worsening patients' conditions. To address this problem, we present a decision support system that automatically recommends therapeutic protocols for AML patients based on their survival prediction. By minimizing subjectivity and streamlining the decision-making process, the proposed approach can enhance patient outcomes, extending survival time and improving overall quality of life.

Treatment decisions for AML rely heavily on predicting the patients' response and clinical outcomes, primarily based on cytogenetic factors (Estey, 2019). However, significant heterogeneity within the same risk groups results in diverse outcomes, ranging from rapid decease to unexpected remission (Döhner et al., 2010).

Chemotherapy has been the established standard therapy since the mid-1970's, but its effectiveness in terms of survival rates has been limited (Bennett et al., 1976). Recent advancements have facilitated the collection and analysis of extensive data on genetic mutations and gene expressions (The Cancer Genome Atlas Research Network, 2013), leading to novel therapeutic strategies and a more targeted approach to treatment. These have opened up new possibilities for improving the outcomes and overall management of AML patients.

In 1976, a study conducted by an international collaboration of French, American, and British researchers known as the FAB (French-American-British) group introduced a classification system for AML. Based on the analysis of morphological characteristics in the bone marrow and peripheral blood, this classification system aimed to stratify AML patients into distinct subtypes. The FAB classification scheme defined six subtypes (M1, M2, M3, M4, M5, and M6) based on the differentiation and maturation levels of the leukemic cells. This classification system was a fundamental framework for understanding and characterizing AML, contributing to subsequent research and guiding clinical approaches (Bennett et al., 1976).

In 2010, the European Leukemia Net (ELN) introduced a novel risk classification system that considers cytogenetic and molecular information, providing a more comprehensive assessment of disease severity (Döhner et al., 2010). This updated scheme includes four risk categories: favorable, intermediate I, intermediate II, and adverse. While this stratification system offers improved accuracy compared to traditional cytogenetic analysis, it is challenging to implement it during the initial clinical evaluation due to the high costs associated with sample collection and the subsequent molecular analyses required. Nonetheless, this risk classification plays a crucial role in guiding treatment decisions and optimizing patient outcomes in the management of AML.

A significant update to the ELN's guidelines was published in 2017 based on findings regarding AML behavior (Döhner et al., 2017). The updated risk classification grouped patients into three categories (favorable, intermediate, and adverse) and refined the prognostic value of specific genetic mutations. Since then, specialists have commonly used this stratification to support important decisions about the course of each treatment, which can directly impact the patient's quality of life and life expectancy.

In 2022, the ELN updated its risk classification system, incorporating significant changes based on emerging research findings. One notable revision includes the FLT3-ITD gene expression as a key determinant. Patients with high expression levels of this gene, lacking other adverse risk characteristics, are now categorized as intermediate risk. Furthermore, mutations in genes such as BCOR, EZH2, SF3B1, SRSF2, STAG2, and ZRSR2 are now associated with the adverse risk classification (Döhner et al., 2022). These updates reflect new insights regarding the impact of these genetic factors on disease progression and treatment outcomes (The Cancer Genome Atlas Research Network, 2013; Angenendt et al., 2019). The evolving understanding of these molecular characteristics provides valuable information for risk stratification and personalized management of AML patients.

Patients with a favorable risk profile typically exhibit favorable responses to chemotherapy. Conversely, those with an adverse risk profile often display limited responsiveness to standard chemotherapy and may require alternative treatments, such as Hematopoietic stem cell transplantation (The Cancer Genome Atlas Research Network, 2013). However, the therapeutic response of AML patients with an intermediate risk profile remains less clearly defined. The heterogeneous nature of this subgroup makes it challenging to predict their specific treatment outcomes, demanding further research and tailored approaches to optimize their clinical management.

The current risk classifications present challenges due to significant variability within the same risk group. Factors such as age and gender can significantly influence treatment outcomes. For instance, patients under 60 years old tend to respond better to high-dose chemotherapy. In comparison, patients over 60 years old may have a lower tolerance for intense chemotherapy and require alternative palliative therapies (Lagunas-Rangel et al., 2017). However, current risk classifications do not consider age a relevant factor in treatment decision-making. As a result, even among patients classified as having intermediate risk, specialists often rely on additional information, such as results from other tests and analyses, to determine the most appropriate therapy despite limited evidence of efficacy (Döhner et al., 2010). This reliance on supplementary information can delay treatment initiation and worsen the patient's clinical condition. Therefore, there is a need for improved risk stratification models that consider diverse patient characteristics to ensure more precise and timely therapy decisions in AML.

To address these challenges, recent studies have applied machine learning (ML) techniques to predict patient survival and treatment outcomes. By leveraging ML algorithms, researchers aim to automate the prediction of patient response to specific treatments and the likelihood of achieving complete remission. These ML models can handle huge clinical and molecular features to compute predictions, allowing for a more data-driven approach to treatment decision-making. The main goal is to provide valuable insights and assist clinicians in making informed decisions that can optimize patient outcomes and improve the overall management of the disease.

Gal et al. (2019) employed supervised machine learning models to predict complete remission in pediatric patients with AML. They used data extracted from RNA sequencing and clinical information as input features for their models. The k-nearest neighbors algorithm achieved the highest performance among the ML techniques evaluated. Additionally, the authors observed notable differences in gene expression patterns between the pre- and post-treatment periods, suggesting the potential of using gene expression data as predictive markers for treatment response in AML patients.

In a subsequent study, Mosquera Orgueira et al. (2021) employed clinical and genetic data to train a random forest classifier to predict the survival probability of AML patients. The researchers identified patient age and gene expressions of KDM5B and LAPTM4B as the three most influential variables. These findings suggest that combining ML techniques with clinical and molecular data holds significant predictive potential for AML diagnosis and supporting therapeutic decision-making. The study emphasizes the importance of incorporating genetic information into predictive models, as it provides valuable insights into the prognosis and treatment response. Such ML-based approaches offer a promising avenue for enhancing patient management and personalized treatment strategies.

Gerstung et al. (2017) presented a statistical decision support model to predict personalized treatment outcomes for AML patients. The model employs prognostic data available in a knowledge bank and demonstrates the significant impact of clinical and demographic factors, including age and blood cell count, on early death rates, particularly mortality related to treatment. Through the knowledge bank-based model, the authors observed that approximately one-third of the analyzed patients would modify their treatment protocols when comparing the model's recommendations to those of the ELN. This highlights the potential of leveraging comprehensive prognostic data and statistical modeling to enhance treatment decisions and potentially improve patient outcomes. The study underscores the importance of incorporating personalized and data-driven approaches in the management of AML.

In a comprehensive study, Itzykson et al. (2021) proposed a rule-based decision support system that integrates statistical and machine learning techniques to facilitate treatment decision-making for elderly patients diagnosed with AML. The model employs overall survival predictions based on the Kaplan-Meier method and incorporates seven oncogenetic markers (NPM1, FLT3-ITD, DNMT3A, NRAS, ASXL1, KRAS, and TP53) to stratify patients into distinct treatment groups. These groups provide insights into the intensity of treatment required and offer guidance on whether treatment decisions should be strictly adhered to, cautiously analyzed, or entirely discarded. The authors found that their model exhibited more discriminative ability than the 2017 ELN stratification, successfully identifying 30–35% of patients with superior outcomes and accurately censoring the need for hematopoietic stem-cell transplantation in the first remission. Moreover, multivariate logistic regression analysis identified mutations in the NRAS, SETBP1, RUNX1, and ASXL1 genes as independent predictors of poor complete remission rates in non-adverse risk patients. These findings further emphasize the significance of incorporating decision-support tools that consider clinical and genetic data for accurate treatment prediction.

An interpretable model for predicting the survival of AML patients was presented by Almeida et al. (2023). This model leveraged the Explicable Boosting Machines (EBM) technique, and the results emphasized the importance of using genetic data in AML analysis, particularly gene expression data. Furthermore, they highlighted the importance of selecting specific treatment groups for patient survival.

Several recent studies on AML-related diseases have also demonstrated the effectiveness of applying state-of-the-art ML techniques in pattern recognition, risk prediction, and survival prediction. These diseases include acute lymphoblastic leukemia (Fitter et al., 2021), myelodysplastic syndrome (Radhachandran et al., 2021), breast cancer (Kate and Nadig, 2017), prostate cancer (Zolbanin et al., 2015; Rabaan et al., 2022), rectal cancer (Wang et al., 2022), skin cancer (Ahmed et al., 2022), nasopharynx cancer (Jing et al., 2020), pancreatic cancer (Walczak and Velanovich, 2018; Muhammad et al., 2019; Wang et al., 2020), infective endocarditis (Ris et al., 2019), AML in pediatric patients (Hoch et al., 2021), and AML with myelodysplasia-related changes (Yu et al., 2021). The success observed indicates that contemporary ML techniques can automatically uncover meaningful patterns within vast datasets.

In this context, this study presents a decision support system designed to recommend suitable therapeutic protocols automatically for AML patients based on their survival prediction. The primary aim is to mitigate the subjectivity inherent in treatment decisions and reduce the time involved in the decision-making process. Consequently, we can deliver more accurate and reliable treatment recommendations that minimize adverse effects. Our ultimate goal is to improve patient outcomes by extending their survival time and enhancing their overall quality of life.

The decision support system proposed in this work combines supervised models computed by three established machine learning methods commonly employed in the medical field: Random Forests (RF), Support Vector Machines (SVM), and Logistic Regression (LR). This system automatically recommends the best treatments for AML patients based on the automatic prediction of clinical outcome (survival/decease).

Altogether, we have trained nine clinical outcome prediction models using selected attributes from real and public databases composed of (i) clinical data (CLIN), (ii) mutation data (MUT), and (iii) gene expression data (EXP). Then, we combined the best-trained models, one for each combination of the three databases. The ensemble outputs are aggregated to compose a robust final prediction model. Figure 1 summarizes the processes for generating the proposed system, and Figure 2 illustrates the architecture of the resulting ensemble models. In the following, we detail each process involved in designing the method proposed in this work.

The final output encompasses recommending a therapeutic course from distinct treatment groups defined by experts in the field and outlined below. As both the individual models and the committee yield survival predictions, the recommendation rests on selecting the group that optimizes the survival forecast for the AML patient.

In this section, we detail and analyze all the results obtained. First, we present the main findings in selecting and analyzing clinical and genetic data. Then, we reported the results of the individual prediction models and, finally, the performance of the outcome prediction computed by the ensemble.

To guide the selection of therapy protocols for patients with AML, healthcare specialists commonly rely on prognostic evaluations based on treatment response predictions and clinical outcomes. The prevailing ELN risk stratification categorizes patients into favorable, intermediate, and adverse risk groups. However, this classification tends to be conservative, with most patients falling into the intermediate risk category. Consequently, specialists often demand additional examinations, leading to delays in treatment initiation and potentially compromising the patient's clinical condition.

This paper presented a decision support system that automatically recommends appropriate intensity oncology therapies based on the clinical outcome prediction for a given patient. The core of this system is composed of a committee of classifiers trained with clinical data and gene mutation and expression data. The proposed ensemble achieved a high performance close to 0.9 in F1-Score and AUC.

We also conducted an evaluation of individual models trained solely with specific types of data. Among them, the model generated with gene expression data exhibited superior performance and could independently assist healthcare specialists in determining the most suitable treatment for individual patients. For further improvement, specialists could employ the ensemble model incorporating all data types. In cases where genetic data collection is unavailable in the clinical setting, the single model trained solely with clinical data can be employed as an alternative.

The findings presented in this work indicate that we can employ state-of-the-art machine learning techniques to automatically process and analyze large volumes of clinical and gene data. These approaches can effectively support specialists in making well-informed decisions regarding the most suitable and safe therapy for individual patients. By significantly reducing the time required for treatment selection, these techniques can enhance overall patient outcomes, leading to extended survival and improved quality of life for individuals afflicted with the disease.

Despite the promising results presented in this study, it is essential to highlight its main limitations. The amount of public data available and used is restrictive to train more sophisticated and accurate machine-learning models. Furthermore, the data represents the characteristics of a particular regionality. At least 75% of the blood samples are from patients of white race, which may hinder the generalization power of the decision support system across different races.

In future work, we aim to assess the performance of the proposed system in a real-world scenario. Furthermore, we recommend further investigating the genes selected in the feature selection stage. This analysis would provide valuable insights into the biological significance and functional implications of the selected genes, mainly MICALL2, KIAA0141, and SLC29A2, potentially revealing novel biomarkers or therapeutic targets related to AML. Such endeavors would contribute to refining and validating the proposed system, ultimately enhancing its application and impact on clinical decision-making processes.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found in the article/supplementary material.

GC: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Validation, Visualization, Writing—original draft, Writing—review & editing. JA: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Validation, Visualization, Writing—original draft, Writing—review & editing. JM-N: Conceptualization, Data curation, Formal analysis, Funding acquisition, Validation, Writing—original draft, Writing—review & editing. TA: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Supervision, Validation, Writing—original draft, Writing—review & editing.