The importance of using machine learning in healthcare

In this article, we will explain what machine learning (ML) is, what benefits it brings to the healthcare industry, and which ML approaches to learning ML models are used.

What is machine learning in healthcare?

Machine learning is a powerful technology that teaches computers to learn from experience. By using mathematics, ML algorithms learn independently from data without the presence of machine operators. The more data is available for learning, the better ML algorithms perform. In other words, algorithms find patterns in data, learn from them, and make conclusions.

In healthcare, ML algorithms process and analyze large amounts of healthcare data and find insights that are impossible to detect manually. They can also predict outcomes, provide risk assessment, and much more.

ML-based solutions, however, are not meant to replace healthcare professionals. Such tools are designed to assist and to work in conjunction with them.

Whether they are used in drug development or patient treatment, ML solutions are changing healthcare for the better.

How does machine learning work?

To train ML models, several approaches can be used. They differ depending on the data sets provided and extent to which they are pre-processed by humans. Data sets can be unlabeled (raw) or labeled (prepared by data scientists).

Therefore, to choose the right approach, it is important to understand which data will be available, whether data scientists can be engaged to prepare data sets, and what specific problems need to be solved.

Three main approaches to training ML models are used:

• Supervised learning
• Unsupervised learning
• Semi-supervised learning

Supervised learning

Along with gathering data, the supervised learning approach requires a number of preparatory steps so that acquired data can be applied. According to this approach, ML algorithms use provided input data and known responses to do the following:

• Learn patterns between inputs and outputs to predict outcomes from new data.
• Recognize objects and features that enable classifying such objects.

Therefore, supervised learning is mostly used for forecasting and classification issues. In general, supervised learning is considered to be accurate, however there are factors that can affect its accuracy:

• Incorrectly chosen input features can lead to incorrect classification results.
• Incomplete or improbable values used for training affect the accuracy of predictions.

Additionally, this approach requires significant computational and human resources to prepare large amounts of quality data sets.

Despite these limitations, supervised learning is widely used in healthcare. Here are a few examples of supervised learning algorithms and their practical applications:

• Decision Trees can be used as a decision support tool for classifying different types of tumors.
• Support Vector Machine finds its practical application in cancer diagnosis, handwriting recognition, etc.
• Naïve Bayes can be used in recommendation systems and for forecasting radiotherapy outcomes.

Unsupervised learning

In unsupervised learning, ML algorithms use unlabeled input data sets. These algorithms discover patterns or information in raw data without human intervention or given instructions. Unlabeled data for unsupervised learning is obtained through collection and observation.

Due to the absence of human intervention, unsupervised learning algorithms can perform more complicated tasks than supervised learning algorithms: discover virtually all kinds of patterns in data and find features to use for classification.

Typical tasks performed by unsupervised learning algorithms are clustering, dimensionality reduction, and anomaly detection.

In healthcare, unsupervised learning algorithms are applied in areas such as genomic sequence analysis, market research, objective recognition, and correct features (variables) selection for healthcare data pre-processing.

However, when choosing this approach, it is important to remember that the results obtained from unsupervised learning can be unpredictable and need to be carefully interpreted.

Below, we will highlight the distinctions between the approaches and provide recommendations for selecting the best method for you.

Patient risk identification

Patient risk prediction is one of the most promising applications of ML in healthcare. By analyzing large amounts of historical patient data, ML models are trained to predict the state of a patient and the likelihood of disease progression in the future. This allows healthcare providers to identify high-risk patients early on and offer targeted care to them.

Personalized treatment

Treatment protocols provide standardized treatment methods for specific conditions. However, patients may often respond differently to the prescribed treatment. ML techniques enable doctors to find patient-specific treatment options based on input data (lab test results, genetics, nutrition, etc.).

Medical imaging

In healthcare, the interpretation of medical images has been traditionally carried out by radiologists and physicians. However, it can be very challenging due to numerous pathologies and subtle changes that may be difficult for the human eye to detect. ML algorithms can be trained to recognize complex malignancy patterns, even in difficult cases or at early stages of disease. As a result, ML tools are becoming particularly helpful in diagnosing and planning efficient treatment.

Clinical trials

Clinical research involves complex, labor-intensive and expensive processes that are prone to errors and deviations. These errors and deviations can lead to the failure of successful trial implementation and acceptance. ML, however, can improve operational and scientific efficiency of clinical trials by facilitating the selection of trial participants, optimizing data collection procedures, and analyzing results.

By utilizing ML-based solutions, researchers can effectively manage large and heterogeneous data sources, identify hidden patterns, and predict the outcomes of clinical trials.

Robotic surgery

When it comes to the application of ML in healthcare, one of the most impactful areas is robotic surgery. ML can be used to cope with various tasks within the field, including but not limited to:

Robotisation of suturing. Even though suturing is an important part of surgical operations, it can be tedious and exhausting. Automating the suturing process can reduce the duration of surgeries as well as ease the burden on operating doctors.

Surgical workflow modeling. Since operations today are becoming more and more complex, continuous improvement is necessary. One innovative approach is surgical process modeling, which aims to enhance safety and efficiency in today’s sophisticated procedures. With the help of ML techniques, surgical workflows can be recreated and models can be constructed taking into account factors such as practitioner’s expertise and patient-specific information, including health status and clinical history. As a result, ML can be of great help in making reliable and timely surgical decisions.

In addition, surgical procedure models can be used for other purposes beyond actual surgery, such as evaluating and training surgeon skills, analyzing workload, introducing new technologies, and predicting procedure duration.

Predicting epidemics

Up-to-date technologies make it possible to forecast epidemic outbreaks based on analysis of vast amounts of data. For prognostication, spread of epidemics in specific areas, along with climatic, demographic, geographic and other factors, is analyzed. This helps identify areas prone to outbreaks and subsequently manage the potential spread of a disease, especially in areas with little or no healthcare facilities.

Clinical decision support systems (CDSS)

CDSS, relying on ML, can analyze large volumes of data, suggest the next treatment options, identify potential issues, and enhance the efficiency of care teams. Typically, decisions related to diagnosis are made using limited patient data. In a CDSS, data of an individual patient are matched to a computerized clinical knowledge base, allowing for patient-specific assessments or recommendations. The latest advances in the development of CDSSs help clinicians make better informed diagnosis, recommend treatment options, and evaluate potential outcomes.

Patient behavior modification

Factors like smoking or lack of physical activity can easily trigger the development of chronic diseases or worsen their course. These factors can be managed or avoided by maintaining healthy habits, which requires guidance and adherence.

To help patients adjust their behavior and adopt a healthier lifestyle, ML-based solutions can be very useful. ML algorithms use patients’ input data (e.g. from wearables) related to their daily activities, analyze them, identify habits and routines, and create personalized plans with constant monitoring of effectiveness.

Drug discovery

The discovery of a valuable drug requires significant investments of time and resources. Thousands of molecules and their combinations must be tested before the right composition is identified. However, utilizing machine learning algorithms in healthcare can accelerate this process by quickly and accurately predicting molecular properties, allowing for more efficient identification of potential drugs.