Article Figures & Data
Figures
- Figure A1.
The Cynefin domains and its application to knowledge in medicine.6
Tables
What is Artificial Intelligence and Machine Learning? • Artificial Intelligence (AI) is a discipline that studies systems that exhibit behaviors associated with human intelligence, whether embodied solely in computer software, or as embedded software in a physical structure like a robot.
• Machine Learning (ML) is an automated approach to learning patterns from empirical data using training examples, usually large databases, with the development of an ML algorithm that when applied to new examples improve prediction. ML is concerned with identification of patterns in a wide range of data include quantitative databases such as diagnostic images or textual corpora such as in medical records, and personal reporting or personal health records. The challenge that ML addresses is to identify patterns of interest in data sets that encode extremely large numbers of factors and with vast numbers of cases from which generalizations are to be formed. Conversely, it is problematic to generalize from relatively small numbers of known cases to recognized patterns for new possibilities.
• Supervised versus Unsupervised Learning -- It is common to distinguish supervised and unsupervised methods, even though the label “unsupervised” is slightly misleading in that for any method to work, a machine-readable data set must be developed and encoded by a human. However, where methods are explicitly supervised, this tends to mean that there is a designated target feature and values within that feature such that training can be tuned to identify with greater levels of supervision. Error analysis from initial training is fed back into the training process and the models can be progressively refined.
• Deep Learning – Deep learning is a subset of machine learning, which is essentially a neural network with three or more layers. These neural networks attempt to simulate the behavior of the human brain—albeit far from matching its ability—allowing it to “learn” from large amounts of data. Although a neural network with a single layer can still make approximate predictions, additional hidden layers can help to optimize and refine for accuracy. Deep learning drives many AI applications and services that improve automation, performing analytical and physical tasks without human intervention.
Complexity Science • Complex systems have hierarchies of power and control.
• They contain independent agents that interact with other agents are often resistant to top-down control.
• Agents interact with other agents that are above, equal to, or below their position in the hierarchy.
• Knowing the parts of a complex system does not equal understanding the system.
• There are non-linear relationships between efforts to change systems (usually from higher levels in the hierarchy) and the amount of change that actually occurs in the lower levels.
o. A complexity science concept that has moved into popular thought is the tipping point, or the notion that a small change can spread from its own momentum across agents in a system without much external energy spent to drive the change.
• In more resilient complex adaptive systems, the top hierarchies provide information and resources to the lower levels. The agents in the lower levels use these resources that make the most sense in their local environment.
• This top-down/bottom-up information and resource cycle results in emergent properties through a complex system that is unpredictable.
• Complexity principles even apply to an individual human. The human “system” contains many hierarchies that interact from the top-down and bottom-up at the same time. The impact of the external environment on health comes from stress activation, etc., at the lowest level. In clinic, we see the effects of heterogeneity of dynamic interactions within body systems at lower level, and environment elements at the whole person higher levels.
- Table 3.
Excited or Apprehensive—Likely Uses of Artificial Intelligence/Machine Learning from a Complexity Perspective
Primary Care Function Excited/Less Complex Care Apprehensive/More Complex Care AI/ML more likely to help Factors making AI/ML less likely to help Diagnose • Improve physician confidence to diagnose a patient concern likely explained by one diagnosis.
• Possibly improve diagnosis of rare diseases.
• More quickly sifting through and sizing up evidence.
∘ Possibly better than doing your own Google or literature search.
∘ More quickly and more focused searches with less time taken from actual patient care.
• Complex interaction of multiple symptoms and possible diseases including psychosocial and social determinant effects that is too much for the computer to manage.
• Actual diagnostic computer “nuts and bolts” can be nonsensical.
• Diagnoses more accurate in AI/ML demonstrations with well-structured vignettes.
• AI/ML diagnoses not likely to include contextual and relational factors.
Treat specific diseases • Better at judging the chances of what treatment will work best for what patient.
∘ Helping the physician become a better “odds-maker.”
∘ Greater objectivity—more evidentiary basis and less sense of exposure
∘ Readiness to take on complex work that might otherwise have seemed “beyond my bandwidth” when not assisted by AI tools, especially for rarer diseases.
• Noisy and inaccurate data in EHRs.
• May only add real value for rarer diseases.
• AI not successful at improving cancer treatment.
• Using historical treatment data may not reflect newer approaches.
• Many treatment decisions are based on a negotiation with the patient using patient-shared decision making.
• Relationships of data elements in existing databases (EHRs, billing data, and so on) often non-linear with Pareto distributions, which makes mathematical predictions challenging.
Predict future health events • Identify patients at increased risk for hospitalization.
• Predictions may not be actionable.
• Existing models may give similar population estimates, but vary widely for individuals.
• AI/ML predictions may not improve upon existing tools.
Decrease administrative burden • Getting through all those letters and forms with less burden.
• For U.S. physicians mostly, decrease documentation time and burden.
• Patient symptoms in primary care are often vague, and even human scribes do not agree how to document them.
• Distinction between typing spoken words onto a clinic note versus increasing understanding.
Abbreviations: AI, artificial intelligence; ML, machine learning; EHR, electronic health record.
- Table A1.
AI/ML Features and Their Relationship to Complex Adaptive Systems and Primary Care
Complexity in Primary Care AI/ML/Big Data Philosophy A complex system is one in which there are so many interacting parts that it is difficult to predict the behavior of the system based on knowledge of its component parts. People are the glue that binds and maintains the system.7 Complex systems have chaotic and non-linear complex parts as well as linear and complicated features. Artificial intelligence (AI) is intelligence—perceiving, synthesizing, and inferring information—demonstrated by machines, as opposed to intelligence displayed by non-human animals and humans.
No established unifying theory or paradigm has guided AI research for most of its history.The Challenge The General Systems Theory (GST)8 is studying the whole system, to clarify the principles that can be applied to all types of systems at all levels of their nesting in the other systems in terms of their viability in all subject areas of research, given their interaction with each other in real-time and in fuzzy environment, surrounding them.9 How to inform the human in the loop with intelligent reliable information that is relevant? Develop intelligent systems thinking in an autonomous fuzzy control, operate in fuzzy environments under uncertainty, and communicate with humans and other systems in different languages, in the dissimilar domains, where processes, situations and factors of influence on the control object and back; a) cannot be determinate and structured in advance, and b) may be understood relevantly and unambiguously.9–11 Responding to the dynamical systems that vary over time. The human brain can be understood as a complex adaptive system itself that have evolved to enable humans to navigate complex environments12 and dynamical systems. Exploration is ongoing in attempts to understand, emotion, context and human reasoning.13,14 Dynamical systems obey differential equations involving time derivatives. Analytical resolution of such equations or their integration over time through computer simulation may facilitate the prediction of the future behavior of the system. Intelligence Human Intelligence aims to adapt to new environments by utilizing a combination of different cognitive processes. The human brain is analogous and uses its computing power to recognize multiple patterns, diverse memories, interoception14,15 and ability to think to make sense16 of their environment and make decisions. Artificial Intelligence aims to build machines that can mimic human behavior and perform human-like actions. AI-powered machines rely on data and specific instructions fed into the system. They have perfect memory but the memory is constrained by their capacity to analyze and infer from human and other inputs. Environmental and sensory Inputs - theory Human consciousness in an adaptation to a new environmental disturbance. Through the conversion of neural cognitive activity - thoughts - about the state of the outside world into the way that world is experienced through the senses, the thoughts gain the reality that sensory images have.17 Human intelligence (explicit) is the main contributing factor that has given definition to the simulations that are created in machine intelligence. Artificial intelligence depends on the best current theoretical models, input data, and the constraints of the AI machine problem-solving skills. Strengths Human cognition has evolved to adapt to our changing world and navigate our environments.18 Garnering both explicit and implicit knowledge, exhibits the highest degree of evolutionary cortical expansion, supported by receptor diversity and human-accelerated genes underpinning synaptic function. Artificial machine learning is a development from human cognition to address weakness in human cognition – access to information, memory, processing time, etc. Machines are better than humans at processing large amounts of data. This will be most useful in the simple and complicated domains of health care knowledge. Weaknesses Humans have limited explicit memory (e.g., cannot easily use the whole of PubMed on a topic) compared with machine intelligence; Humans have as a group a wide spectrum of intellectual capabilities; however these are influenced by stress, unsupportive environments, distractions, lack of access to a vast body of knowledge that is exponentially growing every day. AI is still in its early stages of development. Training AI systems is a time-consuming process. AI will have all the biases and limitations of those who wish to ensure top-down control via protocol vs guidance. Challenges Many tasks are time consuming. Workforce and remuneration challenges undermine personal communication, interoception, and person-centered care. How to self-organize bottom-up care when practice is driven by top-down constraints and funding formulae? Top-down medical industrial for-profit drivers of the artificial data and information systems may come to dominate the primary care space without truly understanding the human and environmental dynamics. How to self-organize and adapt in the complex and even chaotic domains of practice? The information system should serve rather than dominate clinical care.
