Artificial intelligence is currently transforming many industries and it seems that the healthcare industry is the next one. Recently, we can see an increase in the use of machine learning and deep learning in medicine. Deep learning itself is a subset of machine learning, which is a subset of AI, and it uses layered architecture to algorithmically analyze data. Machine learning is a specific kind of computer programming in a way that it uses an algorithm to apply statistical operations against inputs to convert them into outputs without much interference from humans. Although machine learning models become better and better in their function they still need to be controlled and guided to a certain extent whereas a deep learning model can figure out by itself if a prediction is true or not via neural networks.

In other words, in Machine learning an algorithm requires a human with engineering and domain expertise to transform data into representations understandable for the learning algorithm for it to detect patterns. On the other side, a deep learning algorithm needs less preprocessing of data and is capable to develop its own representations for pattern detection from raw data. Thus far the healthcare industry was skeptical when it comes to deep learning techniques and preferred ‘classic’ machine learning because it is explainable. However, in recent years we can see an increase in the use of deep learning, especially in visually-based tasks.

Eager to learn how to build Deep Learning systems using Tensorflow 2 and Python? Get the ebook here!

Currently, there is an unprecedented amount of data that is flowing into the healthcare system but the majority that data is inconsistent – often not labeled properly and feature selection is subjective. That kind of data needs to be properly analyzed and understood. Due to its architecture, the deep learning network can accept that kind of data. The algorithm itself is capable of filtering and processing tasks on data which would have to be handled by programmers if other machine learning techniques would be used. As the data flows through its layers, deep learning methods implicitly learn the features, and gradually useful data points become recognizable – a model is more and more accurate as it processes more data by learning from previous results.

Another convenient feature of deep learning models is their ability to run on specialized computing hardware which is quickly increasing its processing speed – the hardware, especially in the cloud services is more and more powerful. Put differently, a dramatic increase in deep learning algorithms usage is due to the dramatic increase in computational power and the unprecedented amount of available data sets. 

Machine Learning and Deep Learning models learn in two ways:

  1. Supervised learning – Physicians teach an algorithm something, i.e. they label or select a set of diagnoses based on particular clinical signs. 
  2. Unsupervised learning – Identifying data patterns without predictions, i.e. in the early stage of drug discovery where an algorithm is supposed to make its own assumptions.

Supervised learning is more commonly used – the datasets are composed of input data points (i.e. X-ray of the hip) and corresponding output data labels (i.e. fracture/not fracture). This tweet explains the best why supervised learning is more popular than unsupervised learning:

Commonly Used Algorithms

There are many algorithms used in healthcare but the most common ones are:

Support Vector Machines

This is a standard machine learning algorithm that uses supervised learning methods for classification, regression, and detection of outliers. They are used for protein classification, image segmentation and text categorization.

Artificial Neural Networks

A group of deep learning algorithms inspired by the nervous system. More precisely, they are inspired by a neuron organization in animal brains. They consist of units – artificial neurons that receive a signal from the previous layer, process it and send it to the next layer. These networks can learn by “looking” at the examples and without direct programming by humans. Two types of neural networks that are commonly used in the field of medicine – Convolutional neural network and the Recurrent neural network.

Convolutional neural network

It is a feed-forward neural network, a deep learning algorithm that takes in an input image, assigns weights and biases to various features of the input image and after doing that becomes capable to distinguish images one from the other. With enough training its able to “learn” characteristics and filters which would have to be manually programmed in primitive AI methods.

Recurrent neural network

The basic difference between Convolutional and Recurrent neural network is that Convolutional neural network can ingest only a fixed size input thus generating fixed-size output whereas Recurrent neural network is more complex and can ingest arbitrary inputs thus generating arbitrary output data sizes although it does require much more input data than the Convolutional neural network.

Artificial neural networks are used in the fields of computer vision (image recognition, medical image analysis, and feature detection), natural language processing (semantic parsing, speech recognition, sentence modeling, search query retrieval, and text generation), and in early phases of drug discovery (filtering out potentially useful substances and prediction of their medical benefits).

Logistic Regression

It is a machine learning algorithm that uses regression to predict the state of a categorical dependent variable by using predictor variables. It is used for classifications and prediction of an event probability such as a disease risk assessment, a function that assists physicians in medical decisions.

Discriminant Analysis

This is a machine learning algorithm that assesses the adequacy of object classification and also for assigning objects to one group among many groups. Currently, it is used in Electronic health record management and for the early detection of mental health disorders.

Random Forest

It is a machine learning method that constructs multiple decision trees at training time to perform classification and regression. It overcomes the decision trees’ problem of overfitting. It is used currently for ECG and MRI analysis as well as for disease risk prediction based on previous medical history.

Linear Regression

It is a linear statistical and machine learning method for modeling and determining the correlation between dependent and independent variables. In other words, it determines whether one variable is associated with another. It is currently used for disease prediction based on risk factors.

Naïve Bayes

One of the most efficient machine learning algorithms for classification. It is based on Bayes’ theorem with the assumption of strong independence between the observed features. It is currently used for medical data classification and disease prediction.

k-Nearest neighbor

It is a non-parametric machine learning method, a straightforward classifier where the classification of the input data samples is based on the class of their nearest neighbor. Currently, it is used for classification and regression in data mining and heart disease classification.


Although AI transformation of healthcare is imminent and undeniable it does have a few challenges that need to be resolved. One of the biggest challenges is the fact that for proper output, an AI algorithm needs to have a proper input (a huge amount of properly labeled data) and that is difficult to obtain in the current healthcare system.  The fact that AI algorithms require more and more computing power, in other words, bigger data sets require more powerful hardware. 

One more issue that needs to be addressed is that AI models can be inaccurate due to convergence and overfitting, something that will certainly have to be solved in the following years.  And we have to mention one of the biggest concerns in today’s world – the information privacy. AI data sets contain an enormous amount of personal information about patients. Information is money, and a lot of information is a lot of money, a fact that we are being reminded of every time a data breach occurs. Without a solution to that problem, we cannot expect full integration of AI into the healthcare system.


Here we listed only the most frequently used algorithms and addressed the most important issues regarding AI implementation into the healthcare industry. The number of algorithms used and the number of medical fields in which algorithms are performed is continually increasingThe challenges will be solved and it’s expected that in the next 5-10 years AI will be fully integrated into the large areas of the health-tech industry. 

By that time, the AI hype will cease and AI will find its proper place as an irreplaceable tool in the healthcare industry. Due to the specific neural network architecture, very complex functions can be learned since deep learning algorithms can analyze a huge amount of data quickly and precisely. That said, healthcare can benefit from deep learning on a large scale in the future because of AI’s successes in predictive tasks and pattern recognition in three fields – image, language and speech processing. 

There are still many questions to be answered but one question is raised more and more frequently – will doctors be replaced by AI? The short answer is – No. The long answer would be – at least not in the near future since currently all of the AI methods belong to narrow AI and we are not near to constructing general AI entities. If that happens then most of the professions and industries we know today will be transformed or extinct. The real scenario is that now and in the future, doctors will have to be educated about artificial intelligence models and applications in the healthcare industry. That way AI will be able to fulfill its role in a proper way – as a powerful tool that helps doctors in their everyday work.

Thank you for reading!

Nemanja Kovacev

Nemanja Kovacev

MD, Ph.D., Orthopedic and Trauma Surgeon, Java and Python developer, Health Tech SME

Nemanja Kovačev is a health tech subject-matter expert with 15 years of combined medical and programming experience. He defended his ph.d. thesis at the University of Novi Sad, Serbia. Participated in national and international medical and IT conferences as a lecturer and a session chair. Loves to write scientific and popular science articles. Currently a full-time programmer and a health tech SME.