Artificial Intelligence (AI) Black Box
The Contrasts between Clinical Decision-Making tools and Physician Clinical Judgment
Becoming a clinician caring for the sick requires a particular interest and diligence. Physicians habitually take pride in what they give and are enthusiastic about taking that extra step to satisfy their Hippocratic oath. A real doctor is willing to put patients before their welfare. Based on that understanding, it takes inscription, perseverance, discipline, and hard work to merit the title of Medical Doctor or any equivalent privilege in our society.
One can only imagine the value of the clinical Judgment of a physician's clinical judgment; when we look back and appreciate the years of education and sleepless nights a physician has to give up to acquire the skill. In technical terms, a value exclusively refers to learning an algorithm to reach a clinical decision. Having noticed that, regrettably, in today’s healthcare monarchy, physicians only get reimbursed for the services they render to patients merely -utilizing the algorithms (in human terms, the skills, and knowledge of how-to), But not for their know-how, clinical judgment strategies, or decision-making processes.
It is hardly surprising that the Big data industry and Artificial intelligence enterprises are already using physician clinical judgment at no cost, while most physicians seldom appreciate its realism.
The public’s inadequate knowledge of information technologies is by far more destructive than insufficient technology itself because it would resemble the scenario that favors operating a machine gun without learning how to use one.
Deep Learning needs Patient and Physician Data and the Clinical Judgment Process.
Deep learning (DL) is also notorious as deep structured learning. It is an element of a more ubiquitous family of machine learning schemes based on artificial neural networks with representation learning.
Every Artificial Intelligence learns by way of the DL. Deep learning as the instrument of acquiring intelligence for AI necessitates hefty amounts of data feeds, including patient information, physician data, and how a physician reached a particular clinical conclusion on a specific patient case. The clinical decision-making process is practically what an algorithm does in the AI realm.
The overwhelming volume, production momentum, multidimensionality, and implied value of available data are frequently simplified and attributed to “big data.” In the world of modern computers, big data analytics transcend the boundaries of understanding the human brain. Parallel to that, advancements in data analytics and computational power render the opportunity to gain new insight and transfer data-provided added value to clinical practice in real time.
What Physicians need to know about Data Technology
Health information and big data have been the crest of a swiftly developing healthcare technology. The development is more fleeting than the healthcare community can keep up with, thus creating a void that promotes makeshift attention and corner in the healthcare space.
Ideally, healthcare professionals, and physicians in particular, in collaboration with the data expert in the dynamic panorama of modern care, should afford expert knowledge. The clinicians must take leadership in each stage of clinical decision support design, from the data phase and algorithm stage to the decision support level. And the conclusion must ensure transparency at the entire product lifecycle, yet most of all, not to be snatched from them without reimbursement. Physician responsibility should further extend into various ethical domains and must uphold the full-fledged accountability of the product contributors.
Only with clinician advocacy can we dutifully launch and lead interprofessional teams, including patients, and espouse innovative analytic technologies to transpose big clinical data into outcomes that benefit patients and physicians while inspired by absolute human values.
Current Status of Big Data in Clinical Practice
Although medical data gathering and analysis used to be the field in the hands of healthcare professionals and physicians, the widespread accessibility of personal health information on an unprecedented scale has overwhelmingly and irrevocably shifted the landscape of contemporary medical practice. The shift has been so sweeping that even patients are starting to take ownership of their health data by exploiting various technologies such as smartwatches or apps. The modem gadgets of today serve as a meaningful vehicle for health data collection. But contrasting what patients and physicians may assume, they are not currently the sole owner of that Big Data. In effect, with or without the use of Fine-prints seduction, tech companies can and will use it in any shape or form they wish without the originator’s consent.
The Art of Clinical Decision-Making and Mastery of Technological Amelioration
The adeptness of interpreting information into the correct diagnosis and rendering the corresponding treatment option is an everyday habit for all physicians. It involves gathering the pertinent data for each patient, blending it with pre-existing knowledge, forming a clinical decision, and initiating the most suitable treatment in line with the patient expectation and need.
A meaningful share of medical education is devoted to mastering how to differentiate relevant from unnecessary to establish the most desirable Judgment plausible conclusively plausible. Yet, the overwhelming quantity, production rate, multidimensionality, and implied value of today’s medical data transcend the boundaries of recognition of the human brain. Thus, deep learning and Artificial intelligence overcome that constraint by thousands, if not millions. Deep neural understanding can shadow the complete knowledge and skills a person attained in ten years, third or tenth of the time. All DL needs a person to feed the Big Data to it without interruption. But then again, the question is whether the Computer and AI can provide empathy.
Recently, a few tech-savvy has entered the space of what we refer to as “Empathic Transference” This group of scientists sturdily has confidence that a computer can be taught to be empathic, something that has always separated the human from the gadget. Although the topic, as cited earlier, is the subject of considerable controversy, nonetheless, the notion by itself is bothersome to the everyday common. After all, a machine does with it has been taught. And we, at present, need every means of knowing what pertains to those lessons.
Clinical Judgment is about Processing Big Data using a Learned Algorithm.
The clinical Judgment applies to the decision-making process, correspondingly called clinical reasoning. It allows clinicians to reach a clinical decision (clinical decision-making) on how to treat a disease in an individual patient contingent on objective findings and collected subjective patient perceptions.
Artificial Intelligence is Medical knowledge (Algorithm) earned utilizing Deep Learning of preceding Physician Judgments.
Artificial Intelligence needs an algorithm to diagnose, just like a physician needs medical knowledge. Advancements in data analytics and computational power grant the opportunity to obtain new insight and transport data along with subsidized value to clinical practice in real time. The before-mentioned systems are “Clinical Decision Support” (CDS).
Clinical Decision Support is broadly referred to as an information system envisioned to aid in clinical decision-making, combining various references of health information such as Electronic Health Records (EHR), laboratory test results, etc. CDS systems come in many forms and capacities, but all are designed to generate clinically relevant upshots based on input data.
Clinical decision support from the data science perspective typically follows a model as simple as an “if-then rule” by building reference values for laboratory measurements or follows a complex prediction model such as AI pointing radiologists to possible incidental findings on an X-ray image or CT-Scan. The equivalent yield of a CDS system ranges from giving the generated prediction as input for a clinical decision, such as automatically-generated early alert scores, to acting upon the Judgment without human intervention. A typical example of the latter scenario to reference is the implantable cardioverter-defibrillator.
The Computer doesn’t go to Medical School.
If the Computer knows better, one would also assume the medical school curricula must soon become obsolete. Then healthcare and medical practice, in particular, will be another skill earned by the data scientists who develop CDS systems. Most of us may feel such a thought is a hypocrite and absurd. Nevertheless, countless don’t grasp it is, indeed, where we are heading and may well be the reality. But- Is it the right path?
There will always be a need to fill the ethics and empathic gap within the Clinical Decision-making arena. And shortly, there is going to be a necessity for human physician factors for the humanoid Computer to shadow learn. There may be little need for such physicians, who will serve as the masterminds or mentors of Artificial Intelligence and deep humanoid learning. Consequently, the skilled data entry and collection jobs by laypersons will soon be in high demand, something that artificial intelligence technologies in other industry domains are laying the foundation for. An example of those cited above- is collecting valuable personal data through social media.
Clinical Decision making is not as simple as building Algorithms.
In real life, two types of individuals partake during the medical decision-making course; one patient and a doctor. Following the recent introduction to Merit-based physician reimbursement, the role of the patient as a subject of disease experience has become even more imperative. That comes along with their role as shared decision-makers. Yet, the part of the physician and medical expert knowledge is another sphere of data science and holds upon various ethical issues.
Ethical attentiveness relating to algorithmic clinical decision-making justifies a profound conversation. Furthermore, in the epoch of CDS, physicians must initiate and guide interprofessional elements, including patients dutifully and espousing novel analytic technologies to translate big data into patient benefits driven by human values.
A well-designed CDS method demands expert knowledge of physicians throughout the entire three stages of AI development, which include Data, algorithms, and Decision Support.
Medical Data Collection
Data is the first point at which expert knowledge of physicians may enter the CDS development process. However, a substantial part of the day-to-day clinical decision procedure is based on unstructured free-text admissions, encompassing patient history and physical examination data, doctors' observations, or nurses' regular notes. Most tech industry leaders prefer the structured or template data format since it is easier to process. Nonetheless, with new machine learning capabilities, unstructured data is becoming more and more convenient, particularly regarding shadow learning of physician clinical judgment.
Clinical Data Preprocessing
Before the data can be used to construct a representation, they must be preprocessed. Preprocessing steps outline the variables from unprocessed data that are beneficial. During preprocessing, the expert knowledge of physicians comes in very handy, as it is the driver of significant variables and values from the information that was mined or acquired. For example, variables associated with diseases must be used to reconstruct data into an understandable format. Because research guidelines and accompanying questionnaires are not routinely applied in the clinical overhaul. The accuracy of algorithms usually improves if low data values and outliers are excluded, but the absence of data can also carry a “value” only a physician can acknowledge.
Algorithm of Clinical Practice
Once the correct data to develop the CDS system is selected, the following stage is to produce a model that describes the relationship between variables and outcomes in the data. This is achieved using an algorithm or a predetermined computational process to procure a prescription from data. Depending on the complexity of the modeling assignment, algorithm design usually comprises a phase of regular training and a degree of model validation. In the training phase, an algorithm model that best fits the data or can make the most reliable predictions on the generated training data. Then later, in the validation stage, examinations are carried out to verify whether the design is precise. The conclusion is universal to all populations. Modeling and algorithm development are not familiar territories for physicians, but their knowledge and input are precious in their development process.
Algorithms relating to machine learning can be subdivided into two categories; supervised learning algorithms that use expert knowledge about outcomes to guide the process and unsupervised learning algorithms that strive to learn data patterns irrespective of the model result.
Input data for supervised learning algorithms necessitate being marked and chosen manually before modeling, as supervised learning methods rely heavily on expert knowledge. On the other hand, unsupervised learning algorithms aim to reveal consistencies in data without being guided by a pre-labeling of the data. The latter, called clustering algorithms, comprises a scope that can discover new data and populations. This strategy is valuable when information on the features needed when partition between patients and controls is not hitherto available. Unsupervised learning is beneficial in finding starting points for further fundamental scientific research. Therefore, this approach is usually used to find novel patterns in the data, such as how physicians think and make clinical decisions about their interactions with others. An advantage is, thus, that it allows for hypothesis-free or agnostic detection of patterns even when expert knowledge on the difference between subgroups is missing. Unsupervised systems still take advantage of expert physician knowledge in the modeling process.
Clinical Modeling Characteristics
Although all models remain particular for a given enigma, developing a CDS system is not a static method. It often includes rounds of major and minor changes of variables involved and algorithmic fine-tuning. These algorithms are termed “self-learning” and are designed to incorporate newly acquired data over time into their modeling processes.
Clinical Decision Support is Multifaceted
Implementation and use of a CDS system involve multiple processes that include anything from exhibiting the algorithm output distinctly, interpreting it by the physicians, to, ultimately, the executed medical decision.
A CDS scheme is not a simple model offering just an output; to the contrary, it involves some level of interpretation. For instance, Risk rates are conveyed by a color plot, indicating risk compared with a standard disease course. Furthermore, model outcomes must be interpreted in specific medical circumstances before the CDS system can provide the tailored CDS and lead to action. A Deep Neural network or DL is advancing further, making the CDS more independent, necessitating transparency and accountability during its lifecycle. Thus, this step is the part requiring physician supervision.
The Technical Rules of Clinical Decision Making
The fundamentals of CDS are speed, the anticipation of information needed, integration into the workflow, or general ease-of-use advice in alerts.
Acceptation of CDS by physicians depends on the extent to which they feel independent in their decision-making. Rather than choosing colors for the user interface, doctors need to be part of the development process, recognizing the valid data, discussing model design, and validation. This may help physicians feel in charge amid forces disrupting everyday clinical practice. A loyal organization with inspiring leadership encourages the involvement of physicians in developing CDS systems and stimulates its transformation.
Medical practice is tempting to some Technologists.
Parallel to the emergence of machine learning, particularly for deep learning applications in CDS operations; it has become too tempting for IT and data experts to build CDS systems, thus pivoting physicians to become merely data diggers for the tech industry. Nevertheless, human intervention is of even more paramount importance in such a disruptive process. Physicians should not only be included in the execution of the CDS system into clinical practice but also be the central part of an interprofessional CDS development team from the start to the close.
Physicians bring personalized clinical decision-making strategies to the CDS table to understand the circumstances in which variables are gathered during routine care.
It is the prevailing notion that most CDS systems do not intend to supersede physicians but are designed to support them. But then again, physicians must realize the definite necessity and willingness to take on such a role.
Physicians have the Responsibility to their Patients especially concerning Artificial Intelligence.
Clinicians must shield their patients against the irresponsible implementation of data-driven technologies. This uniquely holds valid for self-learning algorithms that self-adapt to the patient population without human mediation. The said phenomenon over time can autonomously shift considerably in the form of “algorithmic drift.” Therefore, it is of pre-eminent importance that physicians initiate and guide the development and implementation of CDS in clinical application.
Data scientists enthusiastically embrace contemporary shifts in AI, feign a bold claim, and carry the burden of proof to equip the healthcare system with proper CDS mechanisms. Once physicians can be persuaded of the added benefit of CDS for their patients, they may acknowledge the need and value of data accumulation, interpretation, and curation. This way, they can readily embrace their expanding role and further develop from a doctor who understands best to a doctor who delivers the best care.
Clinical Judgment is Valuable
As doctors are accountable for the decisions they utter during patient care, they would value the transparency of a model’s judgment process and its evolution to the same extent. Whether or not the particular variables are shared with the physicians is controversial, as sharing certain variables may lead to unacceptable side effects. A CDS system that contains an algorithm that is too complex to understand can result in a so-called “Black Box” circumstance, making it difficult or even improbable for a human brain to comprehend how the forecast model operates. This renders the validation of these black-box algorithms extremely important. That is why human clinical Judgment is of utmost value; something that will complement and satisfy variabilities.
The art of clinical decision-making, or how a physician arrives at a treatment remedy, has value. So losing such a valuable asset to 3rd party without remuneration, mainly to be used unconventionally, is not expected to be acceptable what an average physician.
Clinical Judgment or Clinical Decision Support
Notably, the decision of how to counter a CDS system is a mal on ethical considerations concerning when to treat or not to treat are the expertise of human beings rather than that of AI systems. It is not the CDS system’s scope to determine whether the notion of “First Do No Harm” applies to a specific situation as harm and good and quality of life depend on personal Judgment, circumstances, and inclinations of human beings.
Some patients may choose to take a risk that others may choose otherwise. That includes applying a CDS system with a black box algorithm to their distinct circumstance. This way, the cultural differences may indicate the need for locally tweaked systems. People, whether patients or their loved ones, should participate in shared decision-making, tailoring the usage and outcomes of CDS systems to their desires. What is best for the patient depends on more than the yield of a CDS scheme.
Wake-up call for Doctors
Human beings, doctors, in particular, are creatures of habit. The latter tends to create rules and cultures in their medical practices and hire people who align with that culture. Physicians typically strive with the change, especially when what’s expected are comprehensive mindset shifts, such as moving from fee-based care to value-based medical practice; or dealing with the increasing demands of patients. Artificial Intelligence, Deep Learning, Data Science, and mining evolutions are only a few examples of disruptions requiring mindset-shift from the physician community. Therefore, such a shift better be today than tomorrow, with the hands of physicians and without corporate meddlesome.
