AI in healthcare: from disease to society

The patient, the hospital, the society

Artificial Intelligence certainly has an impact on every industry and aspect of our daily lives. One of the hottest topics is the application of AI in healthcare though still mostly in a retrospective (in silico) validation scenario and not on actual large-scale randomized trials. While we work on achieving General Artificial Intelligence (or at least as general as a whole knowledge area), the way we have been thinking about AI in healthcare is by addressing very specific tasks, one at a time. In the meantime, that’s good enough to cause a major impact in the area while we discover how to learn more holistic representations (either deep or engineered) of the human body.

As mentioned by Eric Topol on his book “Deep Medicine” [1], AI has the potential to make medicine human again, to empower doctors by building better information aggregation tools, better pattern detection techniques, better ways to register information from the visits so they can focus on their actual goal: caring about patients. In order to achieve this ultimate goal, we need to move from our very narrow approach of thinking about patients at an exam-level (e.g., predicting if a patient has cancer from a single image) to a wider approach, a patient-level approach that relies on the whole spectrum of data we have about each person. At the same time, we need to move from populational-driven findings to personalized medicine, where a person is no longer within normal limits (WNL) but on his personalized risk patterns.

Three levels of granularity

While our main focus nowadays is on targeting patient-level, or as I prefer to say, disease-level applications, there is a widely unexplored range of applications we could be tackling in AI in other levels of healthcare. Namely, there is an equivalent (if not higher) potential for optimization at a hospital (i.e., healthcare facility) level and at a government (i.e., policy-making and research) levels. So, while addressing topics such as cardiovascular diseases will prevent the death of 18 million patients per year worldwide [2], improving hospital management would directly impact the 3.5 billion people that have access to healthcare [3], and addressing global policy-making challenges will potentially impact the whole world population.

As you might be thinking, tackling exam-level issues is way more actionable from an individual/company perspective. This can be easily corroborated by inspecting the program of the top conferences in Machine Learning in Medicine (ISBI, MICCAI, EMBC, MIDL) whose sessions and papers are mostly in this direction. Why? It is easier to measure the impact, to move from decisions to actions, to get standardized data, to get to something usable. It’s easier and has a higher impact in the short term. Period.

So, it is expected that most of the initiatives will address this applicational level. However, we need to keep in mind the full picture and keep asking ourselves what can we do at a lower level of the pyramid. While we will cover each one of these levels in detail in future posts in this blog, we wanted to briefly summarize the broad opportunities in each one of them.


AI applied to patients has monopolized the research attention in the last decade. While being the tip of the iceberg, it is a tip with a major impact on our daily lives. AI combined with mobile devices and telemedicine can promote equal access to high-quality diagnoses regardless of the patient’s geographic location.

Most use cases in this area focus on the triage/screening/diagnosis of specific diseases and sensing modalities. Areas with proven success are those where the decision process involves recognizing visual patterns (e.g., dermoscopy, radiology, among others). In this direction, we have been working closely with MobileODT to fight cervical cancer -a preventable major cause of deaths among women- using mobile devices and Deep Learning. In the national/Portuguese landscape, there are multiple references of AI-driven companies for specific health and well-being problems: Adapttech, CardioID, SWORD Health and iLOF.

As we move to a more decentralized data acquisition setting with low-cost devices, tangential use cases come to the table to ensure equal quality of service. For instance, methodologies to automatically validate the compliance of the acquisition protocol (especially relevant with untrained personnel), measuring the quality of the data and interacting with the user to help him improve the procedure. In parallel, as we move from small datasets to massive flows of data, we need tools to summarize the most relevant data, both intra and inter-patient. Here, we can observe for instance the efforts on data summarization for capsule endoscopy which records hours of data from each patient.

Finally, with the appearance of better treatments and procedures to fight each disease, the patient’s adherence to the protocol is key. Thereby, predicting which patients aren’t going to follow the treatment and what actions can we make to improve their adherence is something we need to focus on in AI. See the case of Keheala on how we can impact treatment adherence.


The Machine Learning Spectrum

Do you want to get insights into how we have built such models?

Learn More


It’s estimated that 20% of the healthcare costs are related to administrative tasks [1]. These tasks involve among other things, finding the right billing codes for insurance reimbursement, a task that could be partially automated using RecSys to filter the most relevant codes (we will publish a blog post covering this specific topic soon, stay tuned!).

Resource management is an inexhaustible source of use cases for AI. Here, we can go from forecasting demand for supplies to optimizing resource allocation. Being each minute in a waiting room a life-threatening cause in multiple scenarios, removing each potential bottleneck from the hospital entry door to the patient’s discharge is a must. Matching uncertain/difficult cases with more skilled personnel while ensuring a pedagogic diversity of the cases analyzed by each doctor is also a relevant optimization problem from an AI perspective, where a trade-off exists between the short-term individual patient care and the long-term professional training of the staff.

If we want to transform medicine, we need to focus on improving the overall experience, creating personal relationships between doctors and patients. Therefore, we need algorithms to match the right team of professionals (i.e., physicians, nurses, etc.) with the right patients moving from plain availability to affinity.


As it is the case for hospitals but on a larger scale, governments need to focus on resource allocation and forecasting to match populational needs with quality services. Furthermore, while hospitals tend to focus on daily incoming cases, governments need to look at the whole picture, diving into the appearance of epidemics on optimizing for improving the long-term healthcare system. Governments can use AI to detect surplus or unmatching stocks with demand, proactively optimizing the sharing of resources between the centers in a national healthcare system to adjust to drifts in the requirements.

Regulatory demands and policy-making are typically orchestrated by national organizations. Therefore, it is of their best interest to take advantage of AI to explore better screening programs, resource allocation, compliance verification of devices and facilities, among others. Accelerating these processes without compromising quality, will allow their citizens to get quick access to the latest discoveries in the field.

Information flow

As AI becomes pervasive across multiple decision processes, the potential for model-to-model interaction increases. Thus, we can expect AI predictions to enable other decision workflows in any direction of the pyramid: top-to-bottom, bottom-to-top and sidewise.

  • Top-to-bottom: taking advantage of predictions generated in the upper levels to feed the lower ones. For instance, how can we aggregate patient-level predictions to optimize hospital resource allocation? How can we aggregate information from hospitals to predict global epidemics?
  • Bottom-to-top: considering resource availability and demand forecasting at a hospital-level to optimize the treatment decisions at an individual level.
  • Sidewise: borrowing predictions from other exams and diseases at a patient level allows a more holistic vision of the patient (i.e., moving from diseases to individuals). Similarly, in the middle level of the pyramid, transferring patient/treatment forecasting predictions to support the manufacturing of medical products.

Companies at each level of the pyramid can monetize on the flow of prediction to other levels and other use cases, proven privacy is preserved.

Potential risks

The first challenge is data quality and the gathering process itself. Electronic Health Records (EHR) have been designed to optimize billing (writing down codes) instead of having the full picture of the patient. If you don’t believe me, just think about your last visit to your doctor, answering the very same questions you have been answering appointment after appointment. Therefore, any system that aims to understand patients in a holistic fashion will need to re-think the channel for acquiring relevant data and deploying the AI. EHR -as we know them- are not enough.

Defining the ideal metric to measure the performance of individual-level AI applications is relatively easy, at least at a narrow disease level (i.e., sensitivity, accuracy, positive predictive value) ignoring, of course, less tangible but no less important indicators (e.g., quality of life, emotional status). On the other hand, defining the right metric on the lower levels of the pyramid is easily prone to debate, optimizing efficiency may compromise individuals. We need to understand and merge efforts as a society of what are our Key Performance Indicators in healthcare and how to quantify them.

As AI-driven optimization goes ubiquitous, how much room do we leave for the patients to choose the treatment they want, the side effects they’re willing to deal with on a daily basis? Where do we put the threshold on freedom on matters of public relevance (e.g., vaccination and epidemics)? These are some of the questions we need to ask and answer both as individuals and as a society.

Course, Templates

Data Ignite

Mitigate risks in your AI solutions by mastering our AI methodology.

Learn More

Some notes on the future

AI is changing the world faster than what has been observed before. This makes talking about the future taking a shot in the dark (or with very poor illumination conditions!). However, we have some hunchs of fundamental issues that will/should happen in the near future:

  • Explainability will play a fundamental role in healthcare. Not necessarily as a consequence of regulatory restrictions or to create human-machine trust, but because we need to move from diagnosis to treatment suggestion. It isn’t enough to say healthy/sick. We need tools that help doctors become better at their job, to have mutual feedback between human and machines about why they disagree.
  • Moving from correlations to causal reasoning will be necessary if we want to move from the current loop of masking one symptom with a drug that causes other symptoms, in an unbounded chain of adding drugs to the patient’s daily cocktail.
  • Our tools will evolve in the narrow-global trade-off. From narrow disease/exam (deep learning now!) to a patient-level and from-wide populational-level (medical practice now!) to personalized medicine.

Do you want to further discuss this idea?

Book a meeting with Kelwin Fernandes

Meet Kelwin Learn More


  1. Topol, Eric. Deep medicine: how artificial intelligence can make healthcare human again. Hachette UK, 2019.
  2. WHO. Cardiovascular Diseases. Retrieved from
  3. WHO (2017, December 13). World Bank and WHO: Half the world lacks access to essential health services, 100 million still pushed into extreme poverty because of health expenses. Retrieved from

Like this story?

Subscribe to Our Newsletter

Special offers, latest news and quality content in your inbox once per month.

Signup single post

This field is for validation purposes and should be left unchanged.

Recommended Articles

Can Your Business Optimize AI Predictive Models?

Predictive models are transforming the AI landscape. They can forecast future events, identify past occurrences, and even predict present situations. However, building a successful predictive model is not as simple as it seems. To achieve an effective predictive model, you need to consider three crucial moments: the prediction time, the prediction window, and the data […]

Read More
Is Your Business Ready for Generative AI Risks?

Generative AI is a powerful tool that many companies are rushing to incorporate into their operations. However, it’s crucial to understand the possible risks associated with this technology. In this article, we’ll discuss the top nine risks that could impact your business’s readiness for AI integration. Stay ahead of the curve, and make sure you’re […]

Read More
Can the STAR Framework Streamline Your AI Projects?

As a manager dealing with AI projects, you may often find yourself overwhelmed. The constant addition of promising projects to the backlog can lead to a mounting technical debt within your team, forcing you to neglect the core aspects of your business. Here at NILG.AI, we have a solution for this challenge: the STAR framework. […]

Read More