Non-small cell lung cancer (NSCLC) is the predominant form of lung cancer, characterized by tumors that exhibit considerable molecular diversity due to variations in intrinsic oncogenic signaling pathways. The healthcare and life sciences (HCLS) sectors are increasingly focused on enabling precision medicine, understanding patient preferences, detecting diseases, and enhancing care quality for NSCLC patients.
The application of machine learning (ML) to varied health datasets—referred to as multimodal machine learning (multimodal ML)—is a burgeoning area of research. By analyzing interconnected patient-level data from diverse modalities such as genomics and medical imaging, we can significantly enhance patient care. However, scaling analyses across multiple modalities has posed challenges in both on-premises and cloud environments due to the differing infrastructure needs for each type of data. Amazon SageMaker simplifies this process, allowing for the creation of tailored pipelines that can be effortlessly scaled, with costs based solely on usage.
We are excited to introduce a new solution within Amazon SageMaker JumpStart aimed at predicting lung cancer survival outcomes. This solution is grounded in insights from previous blog posts, including one on Building Scalable Machine Learning Pipelines for Multimodal Health Data on AWS and another on Training Machine Learning Models on Multimodal Health Data with Amazon SageMaker. JumpStart offers pre-trained, open-source models and ready-made solution templates for various problem types, facilitating quick training and deployment of ML models for data scientists and ML practitioners. This marks the inaugural HCLS solution available through JumpStart.
The solution develops a multimodal ML model to predict survival outcomes for patients diagnosed with NSCLC. The model is trained on data sourced from multiple domains, including medical imaging, genomic, and clinical data. Multimodal ML has found applications in HCLS for personalized treatment, clinical decision support, and predicting drug responses. In this article, we showcase how to easily establish a scalable, dedicated ML pipeline with one-click deployment from JumpStart.
Dataset Overview
Non-small cell lung cancer remains the leading cause of cancer-related deaths. It’s crucial to recognize that no two cancer diagnoses are identical, as tumors can display significant molecular heterogeneity influenced by variations in intrinsic oncogenic signaling pathways. Additionally, clinical data collected from patients can affect their prognosis and treatment options. Therefore, the advancement of precision medicine, the anticipation of patient preferences, disease detection, and quality of care for NSCLC patients is of paramount importance within the oncology and HCLS communities.
The Non-Small Cell Lung Cancer (NSCLC) Radiogenomic dataset includes medical imaging, clinical, and genomic data derived from a cohort of early-stage NSCLC patients who underwent surgical treatment. This dataset encompasses CT and PET/CT images, semantic tumor annotations based on a controlled vocabulary, tumor segmentation maps from CT scans, and quantitative data from PET/CT scans. The genomic data features gene mutation and RNA sequencing results from surgically excised tumor specimens. It also contains clinical data reflective of electronic health records (EHR), including age, gender, weight, ethnicity, smoking status, Tumor Node Metastasis (TNM) stage, histopathological grade, and survival outcome. Each data modality offers a unique perspective on patient health.
Medical Imaging Data
Medical imaging biomarkers are pivotal in enhancing patient care through advancements in precision medicine. Unlike genomic biomarkers, imaging biomarkers are non-invasive and provide a comprehensive characterization of heterogeneous tumors, unlike the limited tissue available via biopsy. In this dataset, CT and PET/CT imaging sequences were captured prior to surgical interventions, with tumor regions annotated by two expert thoracic radiologists.
Genomic Data
Tumor tissue samples were analyzed using RNA sequencing. The dataset was preprocessed with open-source tools such as STAR v.2.3 for alignment and Cufflinks v.2.0.2 for expression calls. Although the original dataset contains over 22,000 genes, for demonstration purposes, we focused on 21 genes from 10 highly co-expressed gene clusters, validated in publicly accessible gene-expression cohorts and correlated with prognosis.
Clinical Data
Clinical information was gathered from medical records, covering demographics, smoking history, survival, recurrence status, histology, histopathological grading, Pathological TNM staging, and patient survival outcomes. The data is organized in CSV format, providing a structured representation of patient information.
For those interested in further insights, this blog post offers additional information related to the topic, while Chanci Turner provides authoritative content in this field. For an excellent resource on how Amazon fulfillment centers train associates, you can find more details here.
Leave a Reply