AI competency map: classical AI topics in EP
Education only. Not for diagnosis, treatment or any clinical decision. The examples below describe research, much of it in simulation; none of it is a recommendation for patient care.
Many AI courses organize the field into the same competency areas: machine learning, Bayesian networks and other graphical models, logic, search, planning and game playing, Markov decision processes, reinforcement learning, and AI and society. This page shows how each area connects to cardiac electrophysiology (EP) and mapping, where this course already teaches it, and where it does not yet. Use it to plan a reading path, or to see which parts of an AI curriculum an EP engineer will use first.
At a glance
| Competency area | The idea in one line | Where in this course | Coverage |
|---|---|---|---|
| Machine learning | Learn a mapping from data to labels or values | m35, m37–m43, m47 | Taught in depth |
| Bayesian networks | Reason with probabilities along cause-and-effect links | m35, m38, m41, m48 | Bayes’ rule and simple chains |
| Graphical models | Encode which variables depend on which as a graph | m38; graph structure in m39 | Not yet a topic |
| Logic | Explicit rules that can be read and checked | m38, m44 | Rules and policies |
| Search | Explore states or options in a principled order | m30, m41, m43, m44 | Shortest paths, retrieval, agent loops |
| Markov decision processes | Choose actions in sequence under uncertainty | m41 (as a greedy sequential choice) | Not yet a topic |
| Reinforcement learning fundamentals | Learn a policy from reward by trial and error | Mentioned here only | Not yet a topic |
| Planning and game playing | Plan action sequences; reason against an adversary | m44, m46 | Agent plans; adversarial robustness |
| AI and society | Evidence, fairness, privacy, regulation and oversight | m36, m44–m47 | Taught in depth |
Machine learning
The idea. Fit a model to examples so that it predicts labels or values for new inputs, and measure how well it does on data it has not seen.
Why it matters in EP. Electrograms, ECGs, maps and images are large, labeled or labelable data sets. The course uses supervised learning to annotate electrograms and detect drivers (m35), classify rhythms from the surface ECG (m37), triage wearable alerts (m38), learn on mapping meshes (m39), segment images (m40), locate arrhythmia origins (m41), add physics to learning (m42) and pretrain foundation models (m43).
What an EP engineer needs first. Patient-wise splits, external validation and calibration (m35). A retrospective accuracy is not evidence of clinical benefit (m45).
Bayesian networks
The idea. A directed graph whose nodes are variables and whose arrows say what depends on what, with a probability table at each node. Inference runs Bayes’ rule along the graph.
Why it matters in EP. The simplest Bayesian network has two nodes, a condition and a test result. Its inference is the positive predictive value of m35, and in odds form the likelihood ratios of m48. A chain such as “AF present, the device alerts, a triage filter keeps the alert” is how m38 treats the alerts’ PPV as the prior for the triage step. The Gaussian process in m41 is Bayesian inference over a whole function: it gives an estimate and an uncertainty at every candidate pacing site.
A lesson the graph makes visible. Multiplying the likelihood ratios of two tests assumes they are independent given the condition. Two tests that share a failure cause, such as the same motion artifact, are not, and multiplying their evidence overstates certainty. A Bayesian network draws that shared cause as a node.
Not yet in the course. Learning network structure from data, and inference in networks larger than a chain.
Graphical models
The idea. The family that contains Bayesian networks (directed graphs), Markov random fields (undirected graphs) and hidden Markov models (chains of hidden states that emit observations). The graph says which variables are conditionally independent, which makes inference and learning tractable.
Why it matters in EP. Signals that pass through a fixed sequence of states suit hidden Markov models. One early method delineated the P wave by modeling each beat with a hidden Markov model, then used P-wave features to separate 82 patients with documented AF from 63 controls, with sensitivity and specificity between 65% and 70% (Clavier et al. 2002). Neighboring sites on a map are not independent either; the graph networks of m39 use the same mesh structure but are trained as deterministic networks, not as probabilistic graphical models. Knowing the difference helps when a method claims calibrated uncertainty.
Not yet in the course. Hidden Markov models, Markov random fields and their inference algorithms.
Logic
The idea. Knowledge written as explicit rules, from if-then statements to formal logic, so that a conclusion can be traced to the rules that produced it and every rule can be tested.
Why it matters in EP. Much of EP reasoning is already written as rules. A classic stepwise algorithm for a regular wide-QRS tachycardia applies four criteria in order; in its prospective analysis of 554 tachycardias the four steps together had sensitivity 0.987 and specificity 0.965 for ventricular tachycardia (Brugada et al. 1991). Device and wearable alerts often fire only after a rule such as “k positive windows in a row”, which changes sensitivity and PPV as much as the model does (m38). AI agents run under a policy that says which actions run, which need a person’s approval and which are refused, and guardrails written as rules sit outside the model (m44) (Rebedea et al. 2023).
Why engineers like rules. A rule can be read, reviewed and tested exhaustively; a learned model cannot. Systems in EP often combine both: a learned score and a rule-based gate around it.
Search
The idea. Explore a space of states or options in an order that guarantees, or approximates, the best answer: breadth-first, shortest-path, best-first and heuristic search.
Why it matters in EP. Activation spreading through tissue is a shortest-time problem: fast marching computes arrival times in a single ordered sweep, as Dijkstra’s algorithm does for shortest paths on a graph (m30) (Sethian 1996). Distances on a mapping mesh are found by shortest paths along its edges (m41). Retrieval searches a corpus for the passages a language model should cite (m43), and an agent searches over tool calls (m44). Searching many model designs against one validation set is also search, and it leaks the validation set into the result (m35, m44).
Markov decision processes
The idea. A Markov decision process (MDP) has states, actions, transition probabilities and rewards; a policy maps each state to an action, and the best policy maximizes expected total reward over a sequence of decisions.
Why it matters in EP. Mapping is sequential: each new point, pacing site or image changes what is worth doing next. Active pace-mapping in m41 chooses the next pacing site from a Gaussian process; one Bayesian-optimization method reached 3.9 ± 3.6 mm with 8.0 ± 4.0 sites in simulation (Meisenzahl et al. 2024). That is a one-step-ahead (greedy) answer to a sequential decision problem; an MDP formulation would also weigh how today’s choice changes the value of later ones, and would need a model of transitions and a reward that someone has to choose.
Not yet in the course. The MDP formalism, value functions and the Bellman equations.
Reinforcement learning fundamentals
The idea. Reinforcement learning (RL) learns a policy for an MDP from reward, by trial and error, without being told the right action. Q-learning, for example, learns the value of each action in each state.
Why it matters in EP, and why only in simulation. Trial and error is not acceptable on patients, so RL in EP research runs in simulated tissue. In one study, a Q-learning agent moved over two-dimensional left atrial models built from LGE MRI of AF patients and placed lesions to terminate simulated AF; it terminated AF in 84% of training cases and 72% of test cases, all in simulation (Muizniece et al. 2021). Such results show what a simulator rewards, not what helps patients: the policy is only as good as the simulator (m33, m42) and the reward its designers chose. This course does not teach RL for therapy choices.
Not yet in the course. Q-learning, policy gradients and their pitfalls (reward misspecification, simulator gaps).
Planning and game playing
The idea. Planning finds a sequence of actions that reaches a goal; game playing searches against an opponent, choosing moves that are best against the opponent’s best reply (min-max).
Why it matters in EP. AI agents plan, act and observe in a loop, one tool call at a time, under a step limit and with human approval for any write (m44) (Yao et al. 2023). Game playing’s min-max idea returns in adversarial robustness: training a model to resist small malicious input changes can be written as a min-max problem in which an attacker maximizes the loss and the defender minimizes it (Madry et al. 2017). The fast gradient sign method of m46 is the simplest such attacker (Goodfellow et al. 2014).
AI and society
The idea. Who benefits and who is harmed, what evidence is enough, how data are protected, who is accountable, and how people stay in control.
Why it matters in EP. The course treats this as engineering: study designs and reporting (m45), device regulation (m36), fairness, uncertainty and robustness (m46), de-identification and data governance (m47), and human approval for every write action of an agent (m44). For example, an AI-ECG model trained on a cohort that was 96.2% non-Hispanic white kept AUCs between 0.931 and 0.961 across racial and ethnic subgroups of its test cohort, and its authors recommended reporting performance by race, ethnicity, age and sex (Noseworthy et al. 2020).
Reading paths
- Probabilistic reasoning: m35 (Bayes’ rule) → m48 (likelihood ratios) → m38 (chained alerts) → m41 (Gaussian processes) → m46 (uncertainty).
- Search and sequential decisions: m30 (fast marching) → m41 (shortest paths, active pace-mapping) → m44 (agent loops and plans).
- Rules, robustness and society: m38 (alert rules) → m44 (policies and guardrails) → m46 (adversarial min-max, fairness) → m45 and m47.
What the course does not cover yet
Graphical models beyond Bayes’ rule and simple chains, the MDP formalism, and reinforcement learning are named on this page but not taught in a module. Adding modules for them would change the curriculum and needs the author’s approval.