Recently, scientists have been thinking about how artificial intelligence will benefit the medical field. Ideally, machine learning can leverage the collective experience of almost all clinicians to help a doctor make the wisest decision with millions of similar case experiences as reference. However, in the worst case, the improper use of artificial intelligence can encourage unsafe practices, amplify social prejudice, over-exaggerate results, and even lose the trust of doctors and patients.

Earlier this month, Dr. Alvin Rajkomar and Dr. Jeffrey Dean of Google, and Dr. Isaac Kohane of Harvard Medical School, wrote a blueprint in the New England Journal of Medicine, outlining the promises and pitfalls of machine learning in medical practice.

They believe that artificial intelligence is not just a new tool limited to a particular study or a drug. On the contrary, it is a basic technology that is able to expanding human cognition and to improve every step of medical development. They say that machine learning is not a substitute for a doctor, but rather a contributor using its additional insight to enhance the relationship between the patient and the doctor.

AI in diagnosis and treatment

In the field of artificial intelligence and medicine, diagnosis is one of the most concerned topics.
Even in the early stages of technology development, AI-based diagnostic tools are often better than radiologists and pathologists in discovering potentially fatal lesions on mammograms and in diagnosing skin and retinal diseases. Some artificial intelligence models can resolve the symptoms of mental illness and even suggest treatments.

These initiatives to improve computer diagnostics are due to the latest advances in machine vision and migration learning. Despite the fact that AI typically requires a large number of annotated data sets to "learn", migration learning techniques allow trained AI to quickly learn another similar skill. For example, an algorithm that trained on tens of millions of everyday objects in the standard database ImageNet can be retrained on 100,000 retinal images to diagnose two common causes of vision loss.

In addition, machine learning is ideal for analyzing data collected in daily care to determine what might happen in the future. These systems can bring preventive measures to health problems and reduce medical costs. When it obtains sufficient amount of patient’s health data, AI has been able to establish a more accurate predictive model than using medical raw image data.

The problem is that doctors will have to learn how to gather the necessary information and enter them into the AI prediction engine. These models need to be carefully analyzed to ensure that they are not biased or fabricated.

When proceeding to the next step of treatment, things become more difficult for AI as an AI model trained from treatment data may only reflect the doctor's prescription habits, which may not be the ideal practice. A more useful system must learn from well-planned data to assess the impact of a particular type of treatment on a particular population. This is very difficult. Several recent attempts have found that it is challenging to get expert data to update AI, or tailor this data to local practices. At present, it is still a long way to go before AI is capable of recommending treatment options.

AI in health care reform

The impact of AI on simplifying the workflow of doctors is evident. Intelligent search engines can help identify the necessary patient data, and other technologies such as association input or voice dictation can alleviate the cumbersome process of obtaining medical data, which are already used in doctor’s daily work.

We should not underestimate this particular impact. Doctors are generally busy with paperwork, which takes up valuable time that should be spent with patients. With the assistance of AI, doctors’ workload can be decreased. More importantly, these data can in turn refining the training machine learning model to further optimize patient care and to form a virtuous circle.

Challenges in applying AI into medical field

There are multiple challenges in the coordination between AI and medical communities. The limitations of machine learning should be admitted. For example, if no diverse set of disease data is not collected, the AI model is very likely to be either wrong, biased, or both.

However, this is not a permanent obstacle. AI models are increasingly able to handle unreliable or changing data sets as long as the amount of data is large enough. Although these models are not perfect, they can be further refined with a small annotated set so that researchers and clinicians can identify potential problems through a model.

For example, Google Brain researchers are exploring new ways to enable algorithms to explain their decisions. Making the underlying mechanism clear is becoming more and more important in the clinical environment. Fortunately, the recent AI diagnostic results published in top journals have an intrinsic interpretation mechanism. Although human experts can oversee the development of AI alternatives to reduce false diagnoses, everyone should be aware that zero medical error rates are unrealistic for both humans and machines.

Clinicians and patients who use these systems need to understand their optimal use limits. Neither party should rely too much on machine diagnostics. At present, most of the results we have achieved are limited to models based on historical data sets. In the next few years, the goal is to build forward-looking models that allow clinicians to evaluate in the real world while avoiding the complex legal, privacy, ethical, and regulatory dilemmas associated with acquiring and managing large data sets for artificial intelligence.

Machine learning will never take the place of doctors, but they definitely would be good helpers.

AI is increasingly applied in the pharmaceutical industry. BOC Sciences is working with its partners to actively explore more potentials by offering a wide range of drug discovery services like hit identification, chemical resynthesis, chemical building blocks, chiral resolution and API synthesis using techniques like fragment design, computer aided design, high throughput screening.

Author's Bio: 

BOC Sciences