Pharmaceutical professionals believe artificial intelligence (AI)will be the most disruptive technology in the industry in 2021. As AI and machine learning (ML) become crucial tools for keeping pace in the industry, clinical development is an area that can substantially benefit, delivering significant time and cost efficiencies while providing better, faster insights to inform decision making. However, for patients, these tools provide improved safety practices that lead to better, safer, drugs. Here is how AI/ML can be used to support pharma companies in delivering safer drugs to market.

Overcoming Barriers to Using AI in Clinical Research

Today, AI and ML can be used to support clinical research in numerous ways; including the identification of molecules that hold potential for clinical treatments, finding patient populations that meet specific criteria for inclusion or exclusion, as well as analyzing scans, claims reports, and other healthcare data to identify trends in clinical research and treatments that lead to safer and faster decisions.

However, to take full advantage of the benefits of AI/ML technology, organizations performing clinical trials must first gain access to the tools, expertise, and industry-specific datasets enabling them to build algorithms to fit their specific needs. Healthcare data, unlike purely numerical data pulled from monitoring systems and tools such as IoT or SaaS platforms, is typically unstructured due to the way the data is collected (through doctor visits, and unstructured web sources) and must meet strict security protocols to ensure patient privacy.

To truly leverage AI and ML for clinical research, data must be collected, studied, combined, and protected to make effective healthcare decisions. When clinical researchers collaborate with partners that have both technical and pharmaceutical expertise, they ensure that data is being structured and analyzed in a way that simultaneously reduces risks and improves the quality of clinical research.

The Benefits of AI for Clinical Research

When it comes to research study design, site identification and patient recruitment, and clinical monitoring, AI and ML hold great potential to make clinical trials faster, more efficient, and most importantly: safer.

Study design sets the stage for a clinical research initiative. The cost, efficiency, and potential success of clinical trials rest squarely on the shoulders of the study’s design and plans. AI and ML tools, along with natural language processing (NLP), can analyze large sets of healthcare data to assess and identify primary and secondary endpoints in clinical research design. This ensures that protocols for regulators, payers, and patients are well defined before clinical trials commence. Defining parameters such as these optimize study design by helping to identify ideal research sites and enrollment models. Ultimately, better study design leads to more predictable results, reduced cycle time for protocol development, and a generally more efficient study.

Identifying trials sites and recruiting patients for clinical research is a tougher task than it seems to be at face value. Clinical researchers must identify the area that will provide enough access to patients who meet inclusion and exclusion criteria. As studies become more focused on rarer conditions or specific populations, recruiting participants for clinical trials becomes more difficult, which increases the cost, timeline, and risk of failure for the clinical study if enough patients cannot be recruited for the research. AI and ML tools can support site identification for clinical research by mapping patient populations and proactively targeting sites with the most potential patients that meet inclusion criteria. This enables fewer research sites to meet recruitment requirements and reduce the overall cost of patient recruitment.

Clinical monitoring is a tedious manual process of analyzing site risks of clinical research and determining specific actions to take towards mitigating those risks. Risks in clinical research include recruitment or performance issues, as well as risks to patient safety. AI and ML automate the assessment of risks in the clinical research environment, and provide suggestions based on predictive analytics to better monitor for and prevent risks. Automating this assessment removes the risk of manual error, and decreases the time spent on analyzing clinical research data.

Strategies for Using AI for Clinical Research

During clinical trials, there’s a limited patient population to pull from, as research subjects must meet pre-set parameters for inclusion in the study. On the other hand, as opposed to post-market research, clinical researchers are blessed with vast amounts of information surrounding their patients including what drugs they are taking, their health history, and their current environment.

In addition, because the clinical researcher is working closely with the patient and is well-educated on the drug or product being researched, the researcher is very familiar with all potential variables involved in the clinical trial. To put it simply, clinical trials have a lot of information to analyze, but few patients with whom to conduct the research. Because of this disproportionate ratio of information over patients, every case in a clinical research setting is extremely important to the future of the drug being researched.

The massive amount of patient and drug information available to clinical researchers necessitates the use of NLP tools to analyze and process documents and patient records.NLP can search documents and records for specific terms, phrases, and words that might indicate a problem or risk in the clinical trial. This eliminates the need for manual analysis of clinical trial data – reducing, and in some cases eliminating, the risk of human error while also increasing patient safety. This is especially useful in lengthy clinical trials, for which researchers will need to analyze patient histories and drug results over an extended period of time. Many clinical trials have long document trails and questionnaires that can add up to hundreds of pages of patient data that researchers must analyze.

In a clinical trial, researchers are ultimately trying to determine whether the benefits of a specific treatment outweigh the risks. AI can be especially helpful in clinical trials of high-risk drugs. If a researcher knows that a drug cures or alleviates an illness or condition, but also know that the potential side effects of that drug can have a significant negative impact on the patient, they’ll want to know how to determine if a patient is likely to present those negative side effects. NLP can be used to produce word clouds of potential signals of the negative side effects of a drug that patients would experience.

The only way to do this type of analysis manually is to identify those words using human researchers, then analyze the patient reports to find those words, and group those reports into risk profiles. NLP can automate that entire process and provide insights on risk indicators in patients much more efficiently and safely than human researchers ever could.

Integrating AI & ML with Clinical Research Creates Competitive Results

AI and ML technologies, especially NLP, hold huge promise to support and optimize clinical research. However, that assurance can only be achieved by organizations that have the necessary tools, expertise, and partners to leverage the full benefits of AI and ML. AI and ML solutions support the optimization of clinical research by more efficiently analyzing research data for risks and allowing faster trial planning and research. Those who fail to engage AI and ML for clinical research may find that their competitors are doing so, and as a result, are going to market with new drugs and products faster with higher profits due to decreased research time and safer practices.

Author

Updesh Dosanjh, Practice Leader, Pharmacovigilance Technology Solutions, IQVIA

As Practice Leader for the Technology Solutions business unit of IQVIA, Updesh Dosanjh is responsible for developing the overarching strategy regarding Artificial Intelligence and Machine Learning as it relates to safety and pharmacovigilance. He is focused on the adoption of these innovative technologies and processes that will help optimize pharmacovigilance activities for better, faster results.  Dosanjh has over 25 years of knowledge and experience in the management, development, implementation, and operation of processes and systems within the life sciences and other industries.  Most recently, Dosanjh was with Foresight and joined IQVIA as a result of an acquisition. Over the course of his career, Dosanjh also worked with WCI, Logistics Consulting Partners, Amersys Systems Limited, and FJ Systems. Dosanjh holds a Bachelor’s degree in Materials Science from Manchester University and a Master’s degree in Advanced Manufacturing Systems and Technology from Liverpool University.

The post SEIZING THE OPPORTUNITY TO LEVERAGE AI & ML FOR CLINICAL RESEARCH appeared first on Artificial Intelligence.

Last year, the scientific community built thousands of machine learning models and other artificial intelligence systems to identify COVID-19 on chest X-ray and CT images. Some researchers were sceptical of the results: were the models identifying COVID-19 pathology or were they instead making decisions based on confounders such as arrows and other medically irrelevant features?

To answer this question, two medical students working toward their doctorates in computer science in Su-In Lee’s laboratory at the University of Washington rigorously audited hundreds of machine learning models intended for classifying chest X-rays as COVID-19-positive or COVID-19-negative. Results of their audit are reported in Nature Machine Intelligence.

The domain shift problem

The University of Washington researchers wanted to know whether or not published machine learning (ML) models were generalizable. A generalizable ML model will classify chest X-rays as COVID-19-positive or COVID-19-negative correctly no matter where the chest X-rays came from. A model that isn’t generalizable won’t perform well, for example, when it sees chest X-rays that were acquired at a different hospital.

Computer scientists call this drop in performance domain shift. Machine learning models affected by domain shift pick up on minute, systematic differences between datasets that are stronger and more obvious to the model than subtle indications of COVID-19 infection. These ML models then adopt shortcut learning, training on confounders like arrows and text labels and making spurious associations that emerge even when models are trained and tested on other datasets.

In this way, an ML model that uses shortcut learning will demonstrate domain shift and will not be generalizable, while an ML model that relies on medically relevant features to make decisions is more likely to be generalizable and maintain its performance across datasets.

Auditing, machine learning style

While ML models designed to classify chest X-rays tend to use similar architectures, training methods and optimization schemes, the first hurdle that the University of Washington researchers faced was recreating the published ML models.

“Models can differ in subtle ways…And instead of distributing trained models, researchers give out directions for how they made their models,” says Alex DeGrave, co-first author on the University of Washington study. “There’s a whole range of models that you could end up getting out of that set of directions due to randomness in the [model] training process.”

To reflect possible variations that might emerge during training, co-first authors DeGrave and Joseph Janizek, with their adviser and senior author Su-In Lee, first designed an ML model representative of those introduced in dozens of studies and then made minor adjustments to the representative model. They ultimately created and audited hundreds of models and classified thousands of chest X-rays.

Is it COVID-19 or just an arrow?

After introducing their models to new datasets and observing drops in classification performance indicative of domain shift and shortcut learning, the researchers decided to pinpoint the shortcuts themselves. This is challenging because the decisions made by ML models come from a “black box” – exactly how these models make classification decisions is unknown even to model designers.

DeGrave and Janizek deconstructed this “black box” with saliency maps that highlight regions that a model deems important, applying generative methods that transform images, and by manually editing images. Some saliency maps showed medically relevant areas like the lungs, while others pointed to text or arrows on an image, or to an image’s corners, suggesting that the ML models learned and decided COVID-19 status based on these features rather than pathology.

To validate these results, the researchers applied generative methods to make COVID-19-negative chest X-rays look like COVID-19-positive chest X-rays and vice versa.

“We found that if we went back and fed these [altered] images into the original networks we were auditing, it would typically fool those networks into thinking that they were images from the opposite class,” DeGrave explains. “So that means that the things these generative networks were changing were indeed things that the networks we were auditing looked at.”

The researchers again found that model performance depended upon text markers when they swapped written text on pairs of images (one COVID-19-positive and one COVID-19-negative chest X-ray). The researchers’ experiments also revealed that model architecture had little impact on model performance.

“There’s a lot of focus in the literature, I think, on ‘we have the nicest, most interesting new architecture’. We found that actually has a limited impact…whereas working with the data, and changing the data, collecting better data, had a very sizable impact,” Janizek says.

Building and auditing trustworthy AI systems

The researchers’ results indicate the gravity of shortcut learning. They also point to a need for explainable artificial intelligence, which requires that decisions made by machine learning models be understandable and traceable by humans, going forward.

So, how can researchers build machine learning networks that learn from medically relevant features and are generalizable?

DeGrave and Janizek provide several suggestions. First, researchers should collect data prospectively and with the model’s goal in mind, and datasets should be balanced with good overlap. For example, each institution involved in a study should collect COVID-19-positive and COVID-19-negative data, not one or the other. Second, clinicians should be involved in study design and data collection, and researchers should work with clinicians to identify different kinds of confounders that the ML model might rely on. Third, ML models should be audited before they are applied elsewhere.

These suggestions alone are not enough to overcome shortcut learning, the researchers say, and more research is needed. For now, they hope that this study will spark a broader dialogue about the importance of auditing ML models and the need for explainable artificial intelligence. They also want people to be more aware of the kinds of mistakes machine learning models can make.

“There are methods to explain models and detect shortcuts, there are methods to try to improve models…Researchers need to be really thinking about how all of these methods connect to each other to build not just better methods, but a better ecosystem of methods that connect with each other and make it easy for model developers to build a model that we can trust and rely on,” says Janizek.

The post Machine learning models that detect COVID-19 on chest X-rays are not suitable for clinical use appeared first on Artificial Intelligence.

Clinical nutritionists won’t be left out of the medical AI revolution, as researchers are exploring use cases for augmented diet optimization, food image recognition, risk prediction and diet pattern analysis.

The state of the science is described in a paper published this month in Current Surgery Reports.

Applications for AI and other digital technologies are “still young, [but] there is much promise for growth and disruption in the future,” write multidisciplinary specialists at UCLA Health, San José State University and the Mayo Clinic.

The authors represent expertise in gastroenterology, molecular and biochemical nutrition, anesthesiology and general internal medicine.

Surveying recent research and literature reviews, lead author Berkeley Limketkai, MD, PhD, of UCLA Health and colleagues home in on the four aforementioned use cases. Here are snapshots of their reports on each.

1. Diet optimization. A machine learning model for predicting blood sugar levels after people eat a meal was significantly better at the task than conventional carbohydrate counting, the authors report. The algorithm’s creators used the tool to compose “good” (low glycemic) and “bad” (high glycemic) diets for 26 participants.

“For the prediction arm, 83% of participants had significantly higher post-prandial glycemic response when consuming the ‘bad’ diet than the ‘good’ diet,” Limketkai and colleagues note. … “This technology has since been commercialized with the Day Two mobile application on the front.”

2. Food image recognition. A primary challenge in alerting dieters to likely nutritional values and risks going by photos snapped on smartphones is the sheer limitlessness of possible foods, the authors point out. An early neural-network model developed at UCLA by Limketkai and colleagues achieved impressive performance in training and validating 131 predefined food categories from more than 222,000 curated food images.

“However, in a prospective analysis of real-world food items consumed in the general population, the accuracy plummeted to 0.26 and 0.49, respectfully,” write the authors of the present paper. “Future refinement of AI for food image recognition would, therefore, benefit on training models with a significantly broader diversity of food items that may have to be adapted to specific cultures.”

3. Risk prediction. Machine learning algorithms beat out conventional techniques at predicting 10-year mortality related to cardiovascular disease in a densely layered analysis of the National Health and Nutrition Examination Survey (NHANES) and the National Death Index.

A conventional model based on proportional hazards, which included age, sex, Black race, Hispanic ethnicity, total cholesterol, high-density lipoprotein cholesterol, systolic blood pressure, antihypertensive medication, diabetes, and tobacco use “appeared to significantly overestimate risk,” Limketkai and co-authors comment. “The addition of dietary indices did not change model performance, while the addition of 24-hour diet recall worsened performance. By contrast, the machine learning algorithms had superior performance than all [conventional] models.”

4. Diet pattern analysis. Here Limketkai et al. look at a prospective study of more than 7,500 pregnant women who self-reported dietary intake approximately three months prior to giving birth. Comparing logistic regression with machine learning for predicting adverse pregnancy outcomes, researchers found logistic regression failed to find an association between undesirable outcomes and suboptimal consumption of fruits and vegetables.

Meanwhile the machine learning model “found that the highest fruit or vegetable consumers had lower risk of preterm birth, small-for-gestational-age birth and pre-eclampsia,” which is a pregnancy complication marked by elevated blood pressure and, in cases, organ damage.

Wrapping their discussion, Limketkai and co-authors reiterate that the widening acceptance and use of digital devices as well as AI have:

The post 4 ways machine learning is fixing to finetune clinical nutrition appeared first on Artificial Intelligence.

What just happened? Of all the domains where machine learning is expected to be revolutionary, medicine is perhaps the most universal. In a major new milestone, a drug developed using machine learning is about to enter human trials.

Before a new medicine enters human trials, there is typically three to five years of work behind the scenes, researching causes for diseases and compounds that may help treat them. But working with British AI startup Exscientia, a Japanese drug development company called Sumitomo Dainippon Pharma Co. is about to start phase 1 clinical trials after only 12 months.

The drug in question is DSP-1181, a prospective treatment for obsessive-compulsive disorder (OCD). OCD affects millions of people worldwide, to varying degrees, and can be debilitating in its psychological effects.

Exscientia, based in Oxford, UK, operates an exciting machine learning platform called Centaur Chemist. The platform allegedly takes years off the time required to research new compounds, by combining A.I. techniques with existing knowledge of how medicines interact with the human body.

The benefit of machine learning is that it can happen virtually, and far quicker than scientists are able to work in the real world. The platform can analyze millions of molecular combinations and attempt to identify which may be the safest and most effective in treating a given disease.

Perhaps even more important is the potential savings associated with using machine learning to develop new medicines. Typically, it costs over $1 billion to bring a new drug through from conception to market, with a lot of those costs borne out during the research phases. But taking out years of painstaking research will save both time and money, speeding up development and freeing up resources to develop yet more medicines.

There’s a lot riding on Exscientia and Sumitomo Dainippon’s trial. The first phase is to check how the drug affects the body, and how the body metabolises the drug. So this will not prove the medication’s efficacy.

But if DSP-1181 is shown to be safe, phases two and three can proceed, to see whether the drug can help OCD patients in the real world. And if it does, we’ll witness the dawn of machine learning in medicine.

The post First drug developed using machine learning enters clinical trials appeared first on Artificial Intelligence.

In recent years, cutting-edge computational technologies are increasingly being applied in clinical settings in order to provide higher quality of healthcare. Furthermore, huge amount of biomedical data is continuously acquired and stored from ever accurate medical devices. Conversely, in the scientific research community, there is a growing necessity to develop more optimized methodologies that are able to process big biomedical data and extract discriminating parameters (known as features) for more effective, predictive diagnosis of specific pathologies. To this end, biologically computational intelligence techniques are widely being applied in this field. In particular, recent advances in so called Deep Learning (DL) techniques – including, amongst others, convolutional neural networks, stacked autoencoders, deep reinforcement learning, adversarial learning, transfer learning, meta-learning, end-to-end learning, life-long learning and (semi-/un-) supervised learning with weakly labelled data, graph neural networks etc – have emerged as promising technologies in the clinical and biomedical research domains.

The proposed Special Issue aims to solicit original contributions to demonstrate the potential of DL based methodologies and computational models in challenging the current clinical and healthcare frameworks. Concurrently, to exploit the growing availability of multiple heterogeneous medical Big data (such as neuroimages, electrophysiological time-series, multi-modal biomedical data, electronic health records, etc.), the Special Issue focuses on latest advances in innovative DL approaches to process multi-data sources and develop more accurate, secure and explainable solutions, with potential for deployment in a range of future clinical and healthcare applications.

TOPICS

The topics of interest include, but are not limited to:

• Deep neural networks in clinical and biomedical healthcare applications
• Generative adversarial networks for health data (neuroimages, EEG etc)
• Multi-modal techniques and ensemble architectures
• Automated feature engineering and interpretation of features extracted from biomedical data via deep neural networks
• Transfer learning, meta-learning, end-to-end and deep lifelong learning approaches for improved detection of neuropathologies
• Supervised, unsupervised and semi-supervised learning with (weakly labelled) biomedical data (including electronic health records)
• Deep reinforcement learning and graphical neural networks for electrophysiological signals and /or neuroimages (MRI, fMRI, etc.)
• Deep neural networks for cyber and adversarial attacks in healthcare applications
• New or improved nature-inspired optimization algorithms for DL architectures in biomedical applications
• New hypercomplex deep learning models for 3D and multi-modal signals
• Explainable and privacy-assuring deep learning models and architectures

The post Call for Papers: Advances in Deep Learning for Clinical and Healthcare Applications appeared first on Artificial Intelligence.

While artificial intelligence has the potential to address the epidemic of diagnostic errors in healthcare, the industry must overcome the challenges and limitations of these new digital tools.

That’s the contention of a new book on clinical decision support co-authored by John Halamka, MD, president of the Mayo Clinic Platform, and healthcare writer Paul Cerrato.

“Algorithms that take advantage of machine learning, neural networks and a variety of other types of artificial intelligence (AI) can help address many of the shortcomings of human intelligence,” explain Halamka and Cerrato, who make the point that the complexity of medicine now exceeds the capacity of the human mind.

The book’s authors contend that such complexity “requires humility for clinicians with years of experience successfully diagnosing patients’ ills to admit that they may be missing as many disorders as they catch.”

Nonetheless, while AI has the potential to address many of the shortcomings of human intelligence, the authors also outline in the book the criticisms, obstacles and limitations of this technology—including the fact that the evidence to show that it is having an impact on patient outcomes is mixed.

“Among the criticisms discussed is the relative lack of hard scientific evidence supporting some of the latest algorithms and the ‘explainability’ dilemma,” write Halamka and Cerrato. “Most machine learning systems are based on advanced statistics and mind-bending mathematical equations, which have made many clinicians skeptical about their worth.”

They also make the case that “any attempt to reinvent CDS also needs to tackle the outdated paradigm that still serves as the underpinning for most patient care”—namely that the “reductionistic mindset (that) insists that most diseases have a single cause.” In addition, Halamka and Cerrato charge that the “current medical model relies too heavily on a population-based approach to medicine” and that this “one-size-fits-all model is being replaced by a precision medicine approach that takes into account a long list of risk factors.”

Halamka, who previously served as executive director of the Health Technology Exploration Center for Beth Israel Lahey Health in Massachusetts, joined the Mayo Clinic on January 1.

At Mayo, Halamka now leads a portfolio of new digital platform businesses focused on transforming health by leveraging AI, the Internet of Things and an ecosystem of partners. However, he emphasizes in a recent blog post that while his new book is being published during his tenure at Mayo Clinic, “it is not endorsed by Mayo Clinic and represents the personal opinions of Paul and me.”

Halamka and Cerrato also previously co-authored a 2017 book entitled Realizing the Promise of Precision Medicine: The Role of Patient Data, Mobile Technology and Consumer Engagement. However, they contend that their new book on CDS is the “first to be published about platform thinking” and attempts to provide an in-depth look at the emerging technologies that are transforming the way clinicians manage patients.

At the same time, Halamka and Cerrato note that their “enthusiastic take on digital innovation should not give readers the impression that AI will ever replace a competent physician.” Still, they add that there is “little doubt that a competent physician who uses all the tools that AI has to offer will soon replace the competent physician who ignores these tools.”

The post AI can reinvent clinical decision support, but obstacles remain appeared first on Artificial Intelligence.