Goldman Sachs artificial intelligence ecological report: AI can solve the problems of drug research and development, medical insurance control costs and the efficiency of doctors and hospitals
At the end of 2016, the US Goldman Sachs Group released a 99-page heavy artificial intelligence report: “AI, Machine Learning and Data Fuel the Future of Productivity” (AI, Machine Learning and Data Fuel the Future of Productivity) .
The report focuses on artificial intelligence, expounds the ecology and future of artificial intelligence, and describes the impact of artificial intelligence on medical, agricultural, financial, retail, and energy industries. The report believes that the so-called AI refers to the scientific engineering of manufacturing intelligent machines and computer programs that learn and solve problems in the manner of human intelligence. This area includes natural language processing and translation, visual perception and pattern recognition, and decision making. In recent years, the application fields of Machine Learning (ML) and Deep Learning (DL) have expanded rapidly, and data, faster hardware, and better algorithms are the three cornerstones of the advancement of artificial intelligence. In the following, Arterial Network (WeChat: vcbeat) excerpts the report on the impact of artificial intelligence on the medical field, giving you a glimpse of the future development direction of medical treatment. The report pointed out that by 2025, the average annual cost of medical treatment is expected to save $ 54 billion.
Machine learning has broad application prospects in the medical field. The medical industry needs rich and well-defined data sets, as well as the need to supervise patients anytime and anywhere, and there is great variability in medical results. Machine learning can provide the potential for high returns for many of these sub-industries, such as drug discovery, test analysis, treatment optimization, and patient monitoring. With the continuous integration of artificial intelligence and machine learning, people will be expected to achieve significant “de-risk” in the development of new drugs, which will not only save about 26 billion US dollars in research and development costs per year, but also increase the efficiency of global medical information The value of the cost savings exceeds US $ 28 billion per year.
Where are the opportunities?
Drug discovery and development. Combining machine learning with the drug development process has the potential to improve development efficiency. Machine learning can not only speed up the time range, but also increase the probability of success (POS) of drugs reaching the later stage of the trial. David Grainger, a partner at Medicxi Ventures, believes that False Discovery Rate (FDR) is a statistical phenomenon, and avoiding FDR may halve the risk of the later experimental stage. In addition, in the early stages of drug discovery, the existing virtual screening method is called “high-throughput screening”, and it is very susceptible to FDR. If the risk of the Phase 3 trial can be halved, it can save billions of dollars in costs for large pharmaceutical companies, affect their R & D expenditures of more than 90 billion U.S. dollars and bring meaningful returns, enabling them to free up resources In order to find more potential opportunities.
Remarks: Virtual screening (VS) is also called computer screening, that is, before performing biological activity screening, the molecular docking software on the computer is used to simulate the interaction between the target and the candidate drug, and the affinity between the two is calculated. , In order to reduce the actual number of compounds screened, while improving lead compound discovery efficiency.
Although the huge costs associated with late-stage trials often focus on the design elements of clinical trials, we believe that applying AI / ML to optimizing late-stage decisions on selection criteria, scale, and study length can also achieve meaningful efficiency improve.
Doctor / hospital efficiency. Due to regulatory and divisional reasons, the adoption of new technologies has historically been very slow in the US medical system. In addition to meeting the system challenges, the process from drug discovery to the application of new drugs by doctors and clinics to medical practice is often lengthy and non-contiguous.
Data from Transparency Market Research, a US market research consultancy, shows that a series of decree recently incorporated by the US government into the US Recovery and Reinvestment Act has driven rapid growth in areas such as electronic health records The global market is expected to reach approximately US $ 30 billion in 2023. Data aggregation, continuous improvement of data capture technology, and the continuous reduction of independent hospitals have created an unprecedented opportunity for the large-scale use of data. All this will also improve the functions of machine learning algorithms and artificial intelligence to improve speed, reduce costs and increase accuracy in all aspects of the medical field.
London-based Google DeepMind is working with the National Health Service (NHS) to develop an app designed to monitor kidney disease patients, and a platform formerly known as “patient rescue” to support diagnostic decisions . The key to any AI / ML system is massive data, so DeepMind and NHS have reached a data sharing agreement. NHS will provide DeepMind with dynamic new data streams and historical data for training DeepMind algorithms. Only with massive amounts of data is it possible to analyze clinical data in real time. Of course, if DeepMind can effectively obtain patient data at any time, the insights it can provide will go far beyond the scope of kidney disease.
What’s the pain?
Drug discovery and development. One of the important pain points in the medical field is the time and cost of drug discovery and development. According to data from the Tufts Center for the study of Drug Development, it takes an average of about 97 months for a new drug to reach the market from drug discovery to FDA approval. Although the continuous focus on professional technology can help improve the time span, the cost of new drug research and development continues to increase. According to Deloitte data, since 2010, the approved drug development costs of 12 major pharmaceutical companies have increased by 33% to approximately US $ 1.6 billion per year.
R & D returns. The productivity of biopharmaceutical R & D is still a controversial topic. The cost of developing a successful drug continues to increase, but due to unfavorable factors in the reimbursement system, the reduction in the number of patients, and competition among companies, the income return environment for new drug research and development is not optimistic. Although we expect that the R & D returns for 2010–2020 will increase relative to 2000–2010, the changes between the two are actually insignificant. In addition, one of the most important negative factors affecting R & D returns is the failure of R & D products, especially those that have reached the late stage of testing; the cost of these drugs is estimated to account for more than $ 40 billion per year.
Doctor / hospital efficiency. A particular challenge in the medical field remains that the medical practice of doctors clearly lags behind the approval of new drugs and new treatments. Therefore, many machine learning and artificial intelligence experts in the medical field are constantly encouraging major medical service providers to integrate modern machine learning tools into their workflows so that they can make full use of the collected and published massive medical data storage.
Machine learning and artificial intelligence are expected to reduce the time difference between drug discovery and medical practice; at the same time, they can also optimize treatment. For example, from the North American Radiological Society ’s 2009 study of hepatobiliary radiation, it can be seen that 23% of the second opinion will change the diagnosis conclusion, and this is also an area that machine learning companies focusing on medical imaging are expected to solve. In addition, companies that are committed to using machine learning to make disease judgments at the genome level, such as Deep Genomics, are helping suppliers pinpoint their location to provide more effective and targeted treatment.
What are the current methods for developing new drug development business?
Currently, the drug discovery and development business is an extremely long process of research, testing and approval, which can last for more than 10 years. According to the Tuft Drug Development Research Center, it takes an average of 96.8 months for a drug to advance from the first stage to FDA approval. The development of new treatment methods is a unique challenge not only because of the long time it takes, but also because the POS at all stages of the entire development process is very low.
Drug discovery begins with initial goal determination. Once the goal is determined, people usually use high-throughput screening (HTS) to “hit the discovery.” HTS is very expensive. It is automatically completed by robots. By conducting millions of experiments at the same time, it finds the compounds that have the most potential to reach the goal and improves the “hit” probability of drug discovery. The result of the “hit” is optimized to be the guiding compound, and then further optimized in depth to prepare for the preclinical drug development process. When a drug enters the first stage, the whole process usually takes 1–3 years, and its POS is only 20%.
The first stage: focus on safety; healthy volunteers (POS 20%).
The second stage: the focus is on effectiveness; volunteers with a certain disease or health condition (POS 40%).
The third stage: to further collect information about the safety and effectiveness, dosage and drug combination of different groups of people. The number of volunteers is hundreds to thousands (POS 60%)
How does AI / ML work?
In the medical field, there are a wide range of cases where the advantages of machine learning and AI are perfectly utilized. In those cases, decision-making and / or prediction are not driven by human understanding or intuition, but by data and consideration of various influencing factors that are far beyond the scope of human capabilities. Deep learning also shows its unique potential because it can use the knowledge learned in different tasks to improve performance in other tasks.
Reduce discovery failures and increase POS. People invest a lot of capital in huge opportunity costs to explore and research new treatment methods, and in this process, the probability of success (POS) of the first phase of the trial is only about 20%. Therefore, so far, scholars have advocated the use of AI / ML to develop effective and accurate virtual screening methods to replace the expensive and time-consuming high-throughput screening process.
Recently, researchers from Google and Stanford are working on using deep learning to develop virtual screening technologies to replace or enhance the traditional high-throughput screening (HTS) process and increase the speed and success rate of screening. By applying deep learning, researchers are able to share information from numerous experiments across multiple targets. As Bharath Ramsundar et al. Said in a paper related to machine learning:
“ Our experiments show that deep neural networks outperform all other methods … especially because deep neural networks greatly surpass all existing commercial solutions. On many targets, it achieves near-perfect prediction quality, making It is particularly suitable for use as a virtual screening device. In short, deep learning provides the opportunity to establish virtual screening as a standard step in the drug design pipeline. “(Massively Multitask Networks for Drug Discovery, 2015/2/6)
In 2012, Merck hosted a challenge initiated by data science company Kaggle to determine virtual screening statistics. Now, Kaggle has begun to test the application of deep learning and AI, and has cooperated with AI drug discovery startup Atomwise. Atomwise recently used AI technology to analyze and test more than 7,000 existing drugs in less than a day, contributing to the search for Ebola virus treatment options. According to the company’s statistics, this analysis can take months or even years to complete if traditional methods are used.
Improve doctor / hospital efficiency. Improving diagnosis (Enlitic, DeepMind Health), analyzed radiology results (Zebra MedicalVision, Bay Labs), Genomic Medicine (Deep Genomics) and other fields, and even the use of AI treatment of depression, anxiety and PTSD ( HTTP: // Ginger.io) And other aspects, we have seen some early successful examples of applied machine learning. Due to the continuous development of medical data digitization and data aggregation, medical data will become more accessible. This allows AI / ML not only to reduce the costs associated with process tasks, but also to use algorithms to interoperate disjoint data sets in the past to improve medical care itself. Ultimately, because AI / ML can make considerations beyond the capabilities of humans, it can help suppliers diagnose and treat with higher efficiency.
The cost of drug discovery failure. According to our analysis, through the implementation of machine learning and artificial intelligence, people are expected to halve the risks associated with drug development and discovery in the following situations:
The average annual development cost of approved drugs is US $ 1.6 billion, which includes costs associated with failed drugs (Deloitte).
The annual cost of failed drugs is $ 30 billion, and the funds can be evenly distributed to the approved drug group (Deloitte).
In 2015, the FDA reported 60 approved drugs. This means that, taking into account the R & D costs of failed drugs, the cost of each approved drug during the year was approximately US $ 698 million, of which nearly US $ 42 billion was spent on failed drugs. We believe that machine learning and artificial intelligence can halve the risk of new drug research and development: by 2025, the global pharmaceutical industry will save about $ 26 billion annually.
Accelerate the benefits of transitioning to electronic health records. Currently, in the United States alone, the annual salary of medical information technology personnel has reached about 7 billion US dollars. According to BLS data, due to the aging population and the government’s demand for digital transformation, it is expected that the employment prospects of medical information technology personnel will be greatly improved from 2014 to 2024: compared with 7% growth rate of all other occupations, this The occupational growth rate will reach an astonishing 15%, much higher than the average. However, considering that many job tasks in this profession are easily replaced by automation and software, we believe that machine learning and AI may replace almost all such jobs.
BLS believes that the task of medical information technology personnel is to ensure the quality, accuracy, accessibility and safety of patient medical data used for reimbursement and / or research, and at the same time use technology to analyze patient data to improve the quality of medical care and control costs. The increasing application of AI / ML in the medical industry may have a serious impact on this type of occupation. Based on estimates of per capita medical expenditure and global expenditure share, AI / ML is expected to reduce annual costs by more than $ 28 billion globally by 2025.
Who will be disturbed?
In summary, machine learning and artificial intelligence can save the cost of drug discovery and development, improve POS, and increase the efficiency of suppliers and medical facilities. Therefore, they have the potential to significantly change the prospects of large pharmaceutical companies and the entire medical system. . We have reason to believe that in the long run, machine learning and artificial intelligence technologies will surely proliferate, shorten R & D time, reduce the loss of failed drugs, and intensify competition in drug development.
In addition, efficiency gains and automation may cause some confusion for some medical professionals and companies, especially between those who interpret medical results and diagnosis and those who actually deliver care or perform surgery, such as radiologists, providing Opinion experts and administrative or support staff. We believe that this confusion will exist for a long time, because many technologies are still in the early stages of development, and the cost of adopting these technologies may be slightly higher than other improvement mechanisms.
Challenges of adoption
Although there are obvious opportunities for AI / ML in many sub-fields of the medical field, obstacles to technology adoption still exist.
cost. Implementing AI / ML requires the necessary tools and capabilities, but their cost can be very expensive. Especially in the medical industry, medical cost is still the focus of public attention. In order to ensure that the ML algorithm can make good use of data, people need meaningful capital and expertise, and just ensuring that they have sufficient computing power will cost a lot of money.
Interpretability. The algorithm needs to sort out multiple data sets, and this often generates some so-called black boxes. The medical industry, which has been strictly regulated before, may delay the development of AI / ML applications.
Talent. The obstacles to adopting AI / ML technology may also come from the aggregation of talents in related fields. In 2013, Google paid more than US $ 400 million to acquire DeepMind Technologies; according to news reports, there were only about a dozen members of the team. The difficulty of aggregating such a group of high-level talents and the resulting high costs can be prohibitive.
data. Although the US government has promulgated laws to help digitize electronic health records, there are still challenges in transforming a system that generally uses paper into a fully electronic process. In addition, although many organizations have already passed the threshold of “meaningful use”, fragmentation and lack of availability of important patient data still hinder the further development of reforms.
The above points come from the Goldman Sachs artificial intelligence report: “AI, Machine Learning and Data Fuel the Future of Productivity”