Physician vs technology: Who takes the blame?

>>> 9-1-1, we have an emergency…

AI is not just appearing in our homes with the use of Amazon’s Alexa nor just helping us in transportation with self-driving Tesla cars, but also in healthcare. As we shift from all manual based medicine to an AI-human interactive healthcare, the question in this title is one of the many that needs to be addressed (See Blog Post)

I would like to discuss the view of a fellow blogger, Shailin Thomas, in his Harvard Law Bill of Health blog. He argues that since the algorithm has a higher accuracy rate than the average doctor and going with the suggestion of the algorithm is the statistically best option to follow, then it “seems wrong to place blame on the physician”.

He then focuses on the fact that shifting this blame away from the doctor would lead to a decrease in the exercise of medical malpractice laws , which protect patients by providing a way to police their doctors  final diagnosis, and gives two reasons why this could be a good news.  First is that strict malpractice liability laws do not necessarily mean a better result for the patient as suggested by research done at Northwestern University reported by FoxNews. Second, that taking malpractice liability away from the physician would decrease the overspending in healthcare resulting from the practice of defensive medicine  (ordering more diagnostic tests to avoid potential lawsuits).

Although he certainly makes good use of potential positive outcome of lessening medical malpractice laws on physicians, I strongly disagree that tempering with the serious responsibility of an AI-based diagnostic mistake is the way to effect this change.

The doctor is the healthcare professional, not the AI. Regardless of the use of the AI algorithm, at the end of the day, the algorithm continues to remain a tool for the physician to improve the accuracy and delivery of her diagnosis. As pediatric Rahul Parikh mentions in his MIT technology review, “AI can’t replace doctors. But it can make them better”.  After all, AI is not replacing their jobs, but changing them. Take, for example, if you require a certain complex program to do an aspect of your job you had to do manually before. At first you are sceptic of its use, but as time passes, you become familiar with the program. So much that there comes a point where you decide to fully trust it, bypassing any double inspections. But then you make a mistake. You realize the program made it without crossing your mind to second guess. Will your employer fire you, or the program? In healthcare there is seldom firing, usually there is death.

“There is currently no safety body that can oversee or accredit any of these medical AI technologies, and we generally don’t teach people to be safe users of the technology. We keep on unleashing new genies and we need to keep ahead of it.”

-Enrico Coiera, Director of the Centre for Health Informatics at Macquarie University

The healthcare industry is already good at failing to prevent mistakes. In 2016, medical errors were the 3^rd leading cause of death in the U.S only, where only heart disease and cancer exceeded it. There have been many commentators comparing AI like a black box, as it is nearly impossible to know how the deep learning algorithms come up with their conclusions. But healthcare industries can also be classified as a black box, as Matthew Syed references in his Black Box Thinking book, where everyone trust doctors and hospitals and yet they have let more people die each year than in traffic accidents while the process is overall sorrowful but accepting, with limited post hoc oversight .

Industry	Deaths (2016)
Aviation	325
Traffic accidents	40,000
Healthcare industry (preventable)	250, 000

 

If the physician will not be responsible, who will? The first option would be to hold the AI responsible. However, the AI is inanimate where the affected patient would have no compensation and will end extremely unsatisfied. A second option would be to shift the blame to the developers. This might be difficult as the software can be accurate on its initial design and implementation, for example, IBM Watson in clinical medicine where it was designed to compete in Jeopardy. This would also decrease interest in AI development. A third option would be to hold the organization behind the AI responsible. However, if the AI does not have design failures it would be hard for this to work, like holding Tesla responsible for an accident done by a Tesla car user.

To further develop AI implementation in healthcare, the question of responsibility needs to be addressed. But in your case, who do you think should be held responsible?

 

 

Are YOU living in an infectious zoonotic location?

>>> Are you infected?

Humans cannot see when They play hide and seek

Invisible to the eye, but not to Health

And how does she know how to treat the ones she meets? 

They will play with someone else after her slow, incurable death.

As you may know from reading my About Me section, I had previously worked on drug discovery research for tuberculosis (TB), a leading infectious disease that is ranked 10^th in the leading causes of death worldwide. In humans, the main TB bacterial type, or strain,  causing the disease is Mycobacterium tuberculosis, but there exist other TB bacterial types arising from animals that get transmitted to humans such as Mycobacterium bovis, commonly transmitted by infected dairy products, seals, rhinoceros , and elk . This classifies TB as a zoonotic disease, where zoonosis is any infectious disease that can be naturally transmitted between vertebrate animals and humans.

Within the 1, 415 pathogens (causative agents of the disease i.e. viruses, bacteria) known by 2001 to infect humans, 61% were zoonotic. Deadly pandemic diseases that can be found in history such as the 1918 Spanish flu and the 2009 swine flu, as well as modern deadly diseases like Malaria, the Ebola and Zika viruses, Anthrax, and Rabies have been, and are, zoonoses, or have strains that are zoonotic. In terms of treatments, many human cases caused by these zoonoses, especially the ones from TB and Malaria arising from either bacteria or a parasite, have no cure as their pathogenic strains continue becoming increasingly resistant to current treatments.

Since these diseases start in animals and then get transmitted onto humans, imagine if we could predict in advance which animals are going to carry a disease (called reservoirs), and in which geographical location they would arise. This was the focus of the 2015 rodent study done by Barbara A. Han and colleagues (et al.) from Princeton University.

“By combining ecological and biomedical data into a common database, Barbara was able to use machine learning to find patterns that can inform an early warning system for rodent-borne disease outbreaks.”

-John Drake, co-author in Hahn et al. (2015)

Figure 1. Map portraying hotspots for (A) rodent reservoir diversity and (B) predicted geographical location of future hotspots. (source)

Han et al. (2015) used rodent data from PanTHERIA and applied generalized boost regressions, which builds a prediction model by building up an ensemble of weak prediction models/decision trees to categorize the most important variables for predictions. In this case, predicting zoonotic reservoir status and geographical location. They examined intrinsic traits ( postnatal growth rate, relative age to sexual maturity, relative age at first birth, and production) along with ecological and geographical ones where current zoonotic host carriers have come from.

Among the highlights of the results were that:

• Not every rodent has the equal probability of transmitting the disease, only those that mature quickly, reproduce faster, and live in northern temperature areas with low biodiversity
• North America, the Atlantic coast of South America, Europe, Russia, and parts of Central and East Asia have the majority of zoonotic reservoir happening in upper latitudes (Figure 1)
• Predicted future hotspots include China, Kazakhstan, and the Midwestern USA
• Many rodent reservoir hotspots are within the geographical locations were infectious diseases happen most often (both zoonotic and nonzoonotic)
• The majority of etiologic agents infecting rodents are viruses (Figure 2)

Figure 2. Type of pathogens and parasites infecting the rodent species in the wild, where viruses are the major infecting agents. (source)

Overall, these predictions achieved by machine learning shows that it is possible to predict with accuracy the wild species that carry zoonotic infections, and that machine learning can play a key role in improving how we tackle current incurable infectious disease and future emerging ones.

AI + QSAR: helping drug discovery efforts

>>> Welcome to Club QSAR

Many believe that bringing pharmaceuticals to the market is a quick and easy process that require seldom regulation and experimentation. This is not the case. The drug development process is a long, arduous, and costly one further explained using the figure below.In this post, I will be focusing on the drug discovery ( research & development) stage, which focuses on identifying the perfect drug candidate from many molecules able to have the desired therapeutic effect on a biological target of interest ( i.e., a protein).

This drug candidate identification is done by performing many in vitro ( in glass) experiments that although necessary, consume scientists plenty of costly resources and time that could potentially be saved by using computational instead of experimental means.

Quantitative structure-activity relationships ( QSAR) modelling is the main chemistry-informatics approach used to discover small chemical compounds (drug candidates) having the desired activity against a therapeutic target (usually a protein playing a vital role in the disease) while minimizing the likelihood of off target effects which can cause toxicity. Such predictions help prioritize drug discovery experiments reducing work and resources cost. QSAR usually works using ligand-based models where the protein is ignored due to its complex structure (See Blog Post) while only the small molecule is modeled.

In its simplistic form, the measured activities of many small molecules against a single protein is obtained experimentally, then the model from the small molecules specific features ( fingerprints), i.e., count and arrangements of atoms and functional groups within the molecule. Alternatively, the model can learn these fingertips by deriving them from chemical structures using an auto-encoder.

The most promising QSAR methods prior to deep learning were variations of Random Forests (RF) and support vector machine algorithms. That was before Merck, one of the leading biotech companies sponsored a Kaggle competition to examine which machine learning combinations can provide the most efficient solutions to QSAR problems. The winning entry outperformed RF by using an ensemble Gaussian process (GP) regression, where the primary factor were Deep Neural Networks (DNN) (see insight into DNN).

Use of multi-task DNN in QSAR, for example, have improved the single protein approach mentioned, by allowing the analysis of compounds across multiple proteins. Conceptually, it allows to learn from fewer data by using the fact that molecules having similar features behave similarly across multiple proteins.

Deep learning can also solve a key limitation of both single and multitask models, which is that the activity of molecules against proteins most in need of prediction are the hardest to predict because of scarce data sets.

A promising approach is the use of  Deep Convolutional Neural Network (DCNN) using AtomNet to directly model both the molecule and the structure of the protein to predict bioactivity in novel (new) proteins with no experimental biological activity data for drug discovery applications. AtomNet is the first deep neural network made specifically for structure-based binding affinity prediction.

 

Insight into DNN—————————————————————————————————

DNN are a class of deep learning algorithms made of a network composed of “neurons”. A neuron (a) has many inputs reflected as the input arrows and one output (output arrow). Each of the input arrows is associated with a weight wi. An example to understand weights is if we were to train a model to identify pedestrians in an image but these always appeared in the centre of the image, the model would not be able to recognize pedestrians in other positions of the image as each part of the image would be a different weight. The neuron also has an activation function, f(z), and a default bias term b. A row of neurons forms a layer of neuronal network and DNN has several layers(b), where each output neuron produces a prediction for a separate end point (e.g. assay result)

–——————————————————————————————————————————