With the ubiquitous generation of medical data, there is an urge for effective technological implementation beneficial to the medical and health care sectors worldwide. Advances in medical data analysis and computational concepts have interdependently led to an expeditious rise in the utilization of Artificial Intelligence (AI) techniques for various healthcare data analysis applications, such as risk assessment, detection, diagnosis, prognosis, prevention, data management, and therapy response. Most recently, machine learning has attained sufficient advances to fulfill clinical requirements and abide as one of the significant and effective tools in analyzing extremely complex medical data and establishing clinical AI solutions.

Learning Strategies and Frameworks of Machine Learning:

   Many machine learning techniques have been developed, namely supervised, unsupervised, semi-supervised, reinforcement, and deep learning. Several machine learning frameworks with the most popular algorithms that apply to the medical field are discussed below;

   Supervised learning: Supervised learning imparts the compact framework with the strongest guarantees for medical applications by involving function approximation. Support Vector Machine (SVM), Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) are the recent supervised algorithm employed for cancer diagnosis, organ segmentation, and radiotherapy dose prediction.

   Unsupervised learning: The most unsupervised tasks are related to probability density estimation and utilize Autoencoders (AE), dimensionality reduction methods, clustering domain adaptation for medical domain tasks such as classification of patient groups and image reconstruction.

   Reinforcement learning: Reinforcement learning is energized by behaviorist psychology and helpful in healthcare analysis problems. This strategy utilizes q-learning and markov decision algorithms for tumor segmentation, image reconstruction, and treatment planning.

   Semi-supervised learning: semi-supervised learning is a hybrid strategy that recently applied Generative Adversarial Network (GAN) for some recent clinical applications, including tumor classification, organ segmentation, and synthetic image generation.

   Self-supervised learning: Self-supervised learning is the recent hybrid framework.In medical imaging applications,this framework is scarce,but the latest work relies on context restoration as a pretext task for image classification.

   Transfer learning: Transfer learning re-utilize blocks and layers from the pre-trained model with some data for a certain task and fine-tune it to different tasks. This learning model uses inductive, transductive, unsupervised algorithms for radiotherapy toxicity prediction and adaptation to different clinical practices.

   Ensemble learning: Ensemble learning facilitates improving the overall performance and the stability of the model by merging the output of multiple algorithms to conduct the task. Bagging and boosting are the recent ensemble learning algorithms applied for radiotherapy dose distribution, estimation of uncertainty, and stratification of patients.

Prominent Applications of Machine learning in the Medical Domain:

   In the medical field, machine learning expedites very complicated and time-consuming tasks. Contemporary applications of the machine in the medical sector are listed below;

   Prognosis - Currently, machine learning is widely employed for the prognosis of various types of cancer. Machine learning to attain a large and translational impact on disease prognosis, which focuses on the prediction of disease symptoms, risks, survivability, and personalized medicine providence.

   Electronic Health Records (EHRs) - In the diagnosis process, machine learning models are exploited to extract clinical features from the collection of health records. Recently, machine learning has been applied to diagnose diabetes from EHRs.

   Medical Image Analysis - Efficient extraction of information from medical images is obtained using different imaging modalities, performed with machine learning in medical image analysis. The popular medical image analysis tasks include detection, classification, segmentation, retrieval, reconstruction, and image registration. Fully automated intelligent medical image diagnosis systems are the futuristic healthcare application using machine learning.

   Treatment-Oriented Tasks - The latest application of machine learning in treatment are image interpretation and real-time health monitoring. For image interpretation, advanced machine learning models such as CNN, RNN, and Long Short Term Memory are utilized. Real-time monitoring with machine learning is the key concept applicable for continuous health monitoring in wearable devices, IoT sensors, and smartphones.

   Clinical Applications - Most recent clinical advancements in the machine learning paradigm are disease prediction and diagnosis, Computer-Aided Detection (CAD), clinical reinforcement learning, clinical time-series data modeling, natural language processing for clinical text analysis, and clinical speech and audio processing.

Challenges, Innovative Solutions, and Future Aspects of Medical Machine Learning:

   The applicability of machine learning in the practical healthcare system is considerably increasing, and somehow, it faces safety. privacy, ethical, regulatory, and policy challenges to sustain in medicine. Other than these, some of the implementation complexities of medical machine learning are the availability of diverse and good quality data, lack of standardization, and exchange and distribution shifts.

   Innovative solutions for establishing secure, private, and robust machine learning models in medical applications are mentioned below;

   Privacy Preserving Machine Learning concepts: Preserving the privacy of the users in healthcare is the predominant criteria. In machine learning model training and inference, the model does not reveal the additional and essential information from patient medical data. Few of the emerging approaches to address such constraints are,

   Cryptographic approaches are used when machine learning models need encrypted data from multiple parties, and some of the widely used methods are homomorphic encryption, secret sharing, garbled circuits, and secure processors.

   Differential privacy protects private data by adding perturbation into the datasets. Private aggregation, differentially private stochastic gradient descent and exponential noise differential privacy mechanisms are a few of the recent differential privacy applied algorithms.

   Federated learning is the most recent framework to handle data privacy issues in machine learning and applied for medical applications.

   Measures for Adversarial Attacks: The adversarial attacks of machine learning models for medical applications are resolved by modifying models, modifying data, and adding auxiliary Models.

   Solution for Distribution Shifts: To handle the data distribution shift problem, transfer learning and domain adaptation have been used recently. Domain adaptation is a special type of transfer learning which is particularly utilized for medical image analysis applications. Further, different classes of domain adaptation are categorized as supervised, unsupervised, semi-supervised, and self-supervised domain adaptation methods for medical image analysis tasks.

   Interpretable machine learning, machine learning on edge, handling dataset annotation, distributed data management and machine learning, accountable and model-driven machine learning are some of the open research areas of machine learning in the healthcare sector that require more research efforts and investigations in the future.