Artificial intelligence is altering how humans make decisions.
That means we need to have a good understanding of the real world problems in order to instruct AI’s appropriate; we know that ill-defined prompts produce rubbish results.
In my work developing post-agentic AI-cognitives, I find they perform well when instructed in an iterative way, with progressive precision and specificity rather than one-off interrogations. The objective to achieve metacognition, not just answers.
Here I am in conversation with my AI mind. The image is unprompted visualisation.
Prediction is power over the uncertainty of the problems of the emergent future.
Cassis has transitioned from a healthcare and life sciences consultancy to a developer of tools for use by clinicians and pharmaceutical companies and where the compelling challenges, risks, boundary decisions and costs are. There is a growing suite of cognitive applications which in the main are designed to help clinicians reflect critically on their reasoning and in so doing reduce the risk of avoidable medical errors. (You can likely tell my McMaster University Faculty of Health Sciences background here). Many of these have potential in medical education and others embedded in EHRs.
Dr Mike Tremblay
The focus of my career is clinical reasoning and decision making. I have worked in clinical, management and industry settings and on challenges facing healthcare and pharma, I have formal training in mathematical logic (modal and non-standard logics) and cognitive psychology (human reasoning and decision making).
Boundary Logic and Decision-Making Under Uncertainty
My postgraduate work focused on the application of mathematical logic to reasoning under uncertainty, specifically in domains where decisions cannot be meaningfully reduced to discrete or binary choices. The central premise was that many real-world systems do not operate across well demarcated thresholds, but instead function within boundary regions or zones in which multiple actions remain possible, confidence can be low, and small changes in information, interpretation, or institutional context can materially alter outcomes.
Today, this now maps directly onto the core challenges of modern artificial intelligence and human/machine reasoning.
From Discrete Decisions to Probability Fields
Contemporary AI systems are fundamentally probabilistic. However framed (Bayesian inference, deep learning confidence estimation, or causal modelling), they operate over continuous decision surfaces rather than binary truth states. Yet most human and institutional approaches still force these outputs into high versus low risk, approve versus reject, treat versus watchful waiting.
Most consequential errors, biases, and institutional failures occur at the margin and near decision thresholds, where evidence is incomplete, and human judgement, policy, or incentive structures begin to dominate the outcome.
Real World Relevance
In healthcare, most predictive systems are evaluated on accuracy metrics, yet deployed through frameworks that rely on rules. Clinical pathways, triage protocols, HTA decisions and treatment guidelines often present as pass/fail decision ‘gates’, but are in practice driven by human judgements of risk tolerance, evidence quality, clinician experience, and institutional norms.
My experience provides a lens for modelling this. Rather than asking “What is the predicted risk?”, I ask “How close is this case to the point where different actions become equally defensible, and what factors, whether human, institutional, or technical, are likely to determine the final decision?”
This perspective underpins modern approaches to:
- Bias detection in clinical decision-making
- Regulatory and HTA simulation using gated evaluation frameworks
- Personalised adaptive learning systems that move away from population averages toward individual baselines
- Ethical reasoning that distinguish between rule-based compliance and principle-based judgement.
AI Now
There is growing emphasis on uncertainty estimation, out-of-distribution detection, and epistemic humility: that is, the ability of an AI system to recognise when it does not know. These are, in effect, operational forms of boundary logic. Indeed, for many applying clinical problems to LLMs, they worry that the LLM may not admit to not knowing. In that respect, LLMs do model human hubris.
My brief CV
I have provided consultancy since 1997 and now focus on using AI to create decision support tools for healthcare, pharma/life sciences.
Co-founder Elarin Health, USA, commercialising a predictor of falls by frail elderly.
Co-founder Volv Global, Switzerland, focus on AI and patient identification and rare diseases.
Principal, EDS and Kearney.
Co-founder Eden Communications, UK, which launched the world’s first digital and interactive health TV channel, “Living Health”, in partnership with UK broadcasters. The channel received a number of awards for innovation.
Department director, Hamilton Health Sciences, McMaster University, Canada, as head of a department focused on improving the quality of clinical services. The work focused on clinical workflow, skill mix and patient experience.
Sometime research fellow, University of Kent, UK.
Post-graduate tutor, Faculty of Health Sciences, McMaster University in epidemiology. Amazing place to work.
Senior Lecturer, HSMC, University of Birmingham, UK, taught and published on health policy and management decision making and priority setting; advised the European Commission and the Council of Europe and governments. Associate Dean of Medicine, Deputy Director Public Service MBA and Director Masters in Health Quality Management (quantitative methods / operational research).
PhD University of Toronto, applied psychology.
MA, McMaster University, Hamilton, Canada, application of mathematical logic to reasoning under uncertainty, and the absence of discrete choices (boundary logic).