Fixing the basis first: our analysis of how to scale AI in the medical field by first focussing on the basis infrastructure

 

Applications of machine learning and artificial intelligence in the healthcare domain are steadily rising. The healthcare domain offers numerous exciting applications of AI ranging from patient centered use cases to operational processes. In this blog post we will try making an analysis about the current state of the machine learning and artificial intelligence landscape in the healthcare industry. The goal of this blog is getting a clear overview of the trends and developments in this exciting field and at the same time having a criticial eye to to these trends and developments. To be succesful with applying and scaling AI any organization has to focus on the critical fundamentals first before anything else. We focused our attention especially to the state of applied AI in hospitals. Since we are based in The Netherlands, this analysis is mainly focused on the Dutch landscape but we think it can be applicable to other coutries as well, especially in Europe. In this blog post we will share our point of view and thoughts based on the years of experience we have in the AI field. Unfortunately we can not share all the important topics we have observed in this blog post, so please forgive us if we don't discuss certain topics in this post. Despite that, we did our utmost best in identifying the topics and trends that need careful consideration before applying machine learning and artificial intelligence at scale in the healthcare domain.

The level of adoption of AI by Dutch hospitals

Last year we were fortunate enough, despite Covid-19, to talk with many doctors and medical professionals.  We had knowledgeable discussions with them about the adoption of artificial intellgience and machine learning by Dutch hospitals. Already after a few conversations we had to acknowledge that the level of adoption of AI by Dutch hospitals did not really take off yet. This is not surprising since applying AI in the healthcare domain must be done with great consideration and the adoption of AI in healthcare is accompanied by very strict rules compared to other domains. Many medical professionals told us they are keen to know more in which way AI can help them better treating and diagnosing their patients. Opposite to this they don't feel the urgent need to apply AI in their work. Furthermore, many of them also had to admit they actually have no clue about how AI works and which problems it could solve for them and their patients. Many medical professionals we spoke to believed AI was actually more something of the future rather than a technology that could already help them tremendously on a daily basis when applied in the right manner. These findings were neither suprising nor shocking to us but rather they gave us very useful insights into the current level of adoption of AI in the medical field. These conversations made obvious to us that the current level of adoption of AI in the healthcare domain is still in the exploratory phase. Offcourse the level of adoption also differs among hospitals. Larger academic hospitals are already experimenting with getting machine learning and artificial intelligence models into the clinical practice but overall the real level of adoption is still very low. To accelerate the adoption of AI in the medical field we think a few things could be very helpful on the education side. Specialized courses and training programs for medical professionals educating them about the basiscs of artificial intelligence and machine learning could increase the knowledge of these professionals about the advantages AI can have in the clinical practice. By increasing their knowledge about AI, medical professionals could spot interesting use cases for solving medical problems by AI. Since these medical professionals are in the first line of treating and diagnosing patients, they are crucial for the detection of useful AI use cases in the healthcare domain. Educating these medical practioners would really help move the applications of AI in Dutch hospitals forward. If medical practioners would not educate themselves at least about the basics of AI, many exciting and valuable use cases will be missed. Furthermore it could be very useful to develop a platform or large repository full of examples and uses cases of applied AI on medical problems. This platform shoud especially focus on medical practioners. On one hand, papers of machine learning and deep learning models applied to medical problems rise exponentially and on the other hand only a neglective small fraction of these models work their way into the clinical practice. Offcourse this is also due to some models lacking the necessary quality standards to be useful in clinical practice but the tiny fraction of possible useful models finding their way into clinical practice gives room for thought. This brings us back to the first point we made, many medical practioners just don't know what AI can do for them. Hence it is very important to educate medical professionals about AI but don't expect them to become fluent in reading artificial intelligence papers and research applying artificial intelligence and machine learning models to medical problems. Since these papers are full of technical terms beyond the knowledge of medical professionals. Rather we have to find a solution in the form of training program educting them about AI or a learning platform that exposes all use cases of AI relevant to to their specialization in a clear non technical way. These suggestions could contribute to more knowledge among health care professionals and faster adoption of AI in the healthcare domain.

Fixing the basis

Acknowledging applied AI in the medical field is still in the explaratory phase also gives flexibility for doing things right from the beginning. The last couple of years we have done work for clients which didn't have a solid infrastructure to scale their AI adoption on. In the beginning you will not notice this to a great extent but the more you are going to scale AI the more you will get into trouble. Therefore we urge the medical field to really make careful considerations about their AI infrastructure from the beginning. I have spotted some main points which have to be tackled first before scaling AI in the medical field. Fixing the basics right from the beginning is very important and will save a lot of capital and work later on. These are the main points to be taken into account with great consideration in order to lay a solid groundwork for applying and scaling AI in the medical field:

AI infrastructure: D.I.Y or go for the cloud

The discussion of creating and maintaining a technical infrastructure yourself on one hand or turning fully to the cloud on the other hand is always a source of great debate. Opponents of adopting a full cloud infrastructure point to the privacy issues a fully cloud based strategy can have, where proponents mainly preech the benefits cloud adoption has. We are far more on the proponent side of full cloud adoption especially for AI. In our opninion there is no way to go around cloud platforms when you really want to build and scale a solid AI infrastucture. Regularly we are making use of the Google AI platform and we are still amazed by the total spectrum of AI services Google is offering. It is just absolute not feasible for hospitals or health care organizations themselves to create and maintain the spectrum of AI services Google or other cloud providers like AWS or Microsoft have developed. Google's services for example range from quick experimentation services such as Colab notebooks to advanced machine learning and AI pipelines through custom deployments of Kubeflow. They make data storage, data transformations, model training, data validation and serving your models very straightforward and consistent to do. On top of this you can use many out of the box algorithms which speed up your model development, like bayesian hyperparamter optimization procedures to speed up the optimal hyperparameter search of your models. Google is also the creator of the popular deep learning and machine learning library Tensorflow which fits nicely into the services they offer. And last but not least they expose the best hardware to train your machine learning and deep learning models on. You can train your models on GPUs or TPUs from the best hardware manufacturers like NVDIA. 

All of these services make it not only for healthcare organizations but also for organizations in other domains unthinkable to go around the cloud for adopting a scalable and solid AI strategy. Yes, offcourse this will also bring privacy challenges together with it. But large cloud players have improved a lot on this the last years, plus they have no interest in losing clients by not taking privacy issues seriously. Especially through investigations of the European commision on privacy issues faced by these large cloud providers, they have clearly shifted focus to these privacy issues the last years. Furthermore healthcare organizations could work together in this space to leverage their collective bargaining power to negiotiate the best contracts with large cloud providers. The cloud is a very important part if healthcare organizations want to adopt a solid and scalable AI infrastructure. The earlier they take a deep dive into the AI services of cloud providers the sooner they will reap the benefits.

A topic that can become increasingly problematic is the current state of EHR systems. EHR systems were never made to function as a data storage solution for machine learning and deep learning models. They were developed to store medical data. Since EHR systems are still the most critical source of medical data you can not go around these EHR systems either. Scaling AI solutions in your healthcare organization which make use of EHR systems to extract necessary real time input data, increases pressure and load on these EHR systems. When complex table joins or queries have to be executed in order to expose medical data in real time to machine learning and artificial intelligence models for making predictions this will be become a serious problem: the adoption and scalabaility of AI adoption in your healthcare organization will in part become dependent on the scalability of your EHR system. The only way to avoid this problem is by doing a massive data transfer to the same cloud on which your AI infrastructure is running or creating a real time highly scalable datastore of EHR data, preferably in the format of a massive feature store. But this is something we don't see happening soon, so we think for the foreseeing future healthcare organizations will stay dependent on these EHR systems. Healthcare organizations have to discuss these concerns with providers of EHR systems and desire EHR providers will work on making their systems scalable for increasing future loads of AI models. This will be a crucial consideration if you want to take AI adoption seriously for your healthcare organization.

AI development: third parties, in-house data science and software engineering teams or a mix of both

This is a very important consideration healthcare organizations have to take into account. The most logical evolvement of adopting an AI strategy by healthcare organizations consists out of a hybrid mix of in-house devlopment, partnerships with third parties for model development and buying out of the box working models from companies. These considerations are important since we have to avoid artificial intelligence adding to increasing healthcare costs instead of helping to decrease healthcare spending.

The scope of AI is potentially so large that hospitals will not be able to reach the full scale and potential of AI adoption without working together with third parties. It takes a long time for taking an artificial intelligence and machine learning model from an initital idea to production. If you want to bring multiple models to production each year you already need a massive data science and software engineering team. Therefore healthcare organizations have to make careful considerations which use cases to outsource and which use cases to develop in-house. Important considerations you have to take into account are the cost structure of outsourcing models versus in-house development, maintence costs, and lock-in by third parties. As a healthcare organization you could for example set up a cloud AI infrastructure yourself and force third party providers to structure the machine learnining code and additional pipelines in such a way that the code and pipelines could run on your infrastructure. In this way lock-in can be prevented and partnerships can be terminated when agreements are not met. At least we would recommend healthcare organizations to develop or maintain a central model repository. Healthcare organizations also have to take care they are not developing the same components multiple times, analogous to the don't repeat yourself mantra in software engineering. In applied machine learning and artificial intelligence you can also use the same components in multiple projects. If you are working with multiple third parties and they are all individually creating the same components, this would be a waste of resources. These component can be both inherent to your models, pipelines or other parts of your infrastructure.

Sharing platforms for AI and building fundamental algorithms before anything else

This is something we are very excited about. Some fundamental algorithms are used by data scientists before they are really developing a model. A lot of these fundamental algorithms are used in the data preperation and data transformation phase of developing a model. Examples of these fundamental algorithms deal with data augmentations and transformations that have to be applied on your data. In many instances there is not enough medical data to train an algorithm on. As a solution to this you can for example make use of generative algorithms to augment your data. Word embedding are another example. Before you can use natural language as input to a machine learning or AI model you have to transform natural language to word embeddings. Creating useful word embeddings is a extensive project on it's own. It would be wise to create the word embeddings by a specialized team or third party focussing on this problem. By subsequently opening up these word embeddings to third parties or in-house teams working on machine learning models using textual data, these teams or third parties don't have to put in the work anymore to create accurate word embeddings themselves, saving valueable time and resources. This will save a lot of extra work and therefore costs. We can sum up many more fundamental algorithms to show the extent of its importance but we think we made our point clear. Healthcare organizations adopting AI have to put focus on these fundamental algorithms to scale AI wisely in a cost effective way. This is very important and we hope healthcare organizations will see the benefits of focusing on these fundamental algorithms. In the beginning you will have to invest resources into creating and training these models but in the long-term this will save a lot of work and capital.

Besides focusing on these fundamental models we think a platform through which different hospitals can cooperate on machine learning models could also have tremendous value. There are actually two ways to approach such a sharing platform. The first one is by healthcare organizations sharing and bundling their data and subsequently collectivelly paying for model development. But if you want to truly scale AI, collecting data every time a new model is going to be developed will quickly become a massive and unscalable task.

Rather we would go for another choice based on a well known machine learning principle: transfer learning. Transfer learning makes it possible to train a pre-trained model on your own data. For example a hospital different from your own could create a machine learning model trained on their EHR data. Via a shared model platform they could share this pre-trained model with your hospital. You can use this pre-trained model to exponentially scale up the training time of this model for your organization compared when you would have started from a model that was not pre-trained. Many private companies probably use pre-trained models together with data augmentation techniques to make up for the lack of data they have. We think putting focus on a platform that makes model sharing between healthcare organizations feasible would result in great benefits for all healthcare organizations participating in such a platform. Furthermore it would open up models for smaller hospitals which don't have the resources to develop models on their own or spend numerous amounts of money on third party model development.

Concludingly, the adoption of AI in the healthcare domain opens up great and exciting possibilites but careferul considerations have to be made about AI infrasturcture, model development and fundamental concepts shaping the basis for scaling and applying AI in the healthcare domain. We are very excited about the benefits AI could have for patients and healthcare organiazations but they have to fix the basis first before anything else, otherwise they risk getting into trouble later on. With healthcare costs already exploding in many developed countries we have to avoid AI becoming a contributor to increasing healthcare spending instead of decreasing it. We hope this blog post will help medical professionals take careful considerations about applying AI in the healthcare domain. Hopefully this blog post contributes a small amount to the understanding of medical practioners about the challenges we have to solve first before scaling AI in the healthcare field.