Extended Nomological Deductive Reasoning (eNDR) is a novel Explainable AI framework that integrates causal domain knowledge with deductive logic to deliver transparent, trustworthy predictions in dynamic, high-stakes environments. Building on the original NDR model, eNDR handles continuous data, uncertainty, and real-time inference by expressing domain laws as continuous functions, modeling conditions as constraints, and generating explanations through diferentiable reasoning and probabilistic integration. Early results show that eNDR produces human-readable, domain-aligned explanations without sacrificing predictive performance, ofering a pathway toward AI systems that are both accurate and interpretable across domains such as healthcare, finance, and criminal justice.
The field of Explainable AI (XAI) has produced over 200 techniques aimed at improving model transparency. Popular model-agnostic methods like LIME and SHAP ofer post hoc explanations by approximating black-box decision boundaries but often lack domain-specific reasoning, limiting their efectiveness in expert-driven fields.
Inherently interpretable models—such as decision trees and rule-based systems—ofer more understandable outputs but typically sacrifice predictive performance, highlighting the persistent trade-of between accuracy and interpretability in XAI.
A promising alternative is the Nomological Deductive Reasoning (NDR) framework, which balances performance with explanatory depth by integrating deductive logic and causal domain knowledge through the Nomological Deductive Knowledge Representation (NDKR). Inspired by Hempel’s covering law model, NDR aims to provide logically structured, cognitively satisfying explanations.
Originally proposed by Hakizimana and Ledezma Espino, NDR was limited to static, deterministic settings. To address real-world complexity, the extended version—eNDR—introduces probabilistic reasoning, continuous data handling, and real-time inference, making it more applicable to dynamic, high-stakes domains like healthcare, finance, and criminal justice.
This research will explore the following key questions: 1. How can the Nomological Deductive Reasoning (NDR) framework be extended to handle continuous data and domain-specific uncertainty while preserving the interpretability and trustworthiness of the AI model’s explanations? 2. How can the integration of structured knowledge and causal reasoning improve the transparency and explainability of AI systems in high-stakes applications?
By modeling the Nomological Deductive Reasoning approach through the expression of laws as continuous functions; conditions as constraints on data instances; predictions as a function of data instances, laws, and conditions; and explanation as an integral summing over all possible laws and conditions capturing the cumulative efect of how they interact with the data, the NDR framework can handle continuous data and complex real-world scenarios. In addition, using probabilistic settings and considering the prediction task as an optimization process can enhance the interpretability and transparency of NDR-based predictions in domains with complex data and uncertainty. 3.3. Objectives 1. To extend the NDR framework to accommodate continuous data and uncertainty through advanced mathematical methods like calculus and optimization techniques. 2. To test the efectiveness of the extended NDR framework in generating human-comprehensible explanations that are aligned with domain-specific knowledge. 3. To evaluate the performance of AI models using the NDR framework in real-world applications, such as healthcare, finance, or criminal justice.
The research approach is structured around the extension of the Nomological Deductive Reasoning (NDR) framework.The inclusion of probabilistic models and uncertainty quantification within NDR allows the system to account for variations in the input data and make robust, well-grounded predictions even in the face of noise or uncertainty, enhancing both the model’s reliability and its interpretability when making decisions on complex data.
In our initial development of NDR framework, we have assumed the following:
Let represent the set of laws or rules governing a certain real-world domain (e.g., healthcare diagnosis, bank credit score, trafic code for mobility applications, criminal justice, etc.). These laws are formalized as logical statements or principles that provide the foundation for reasoning in the system. Each law ∈ corresponds to a specific rule or law within the system.
Example (in medical settings): • 1: “If a patient has high blood pressure and is over 60 years old, then they are at a high risk of cardiovascular disease.” • 2: “If a treatment is an ACE inhibitor, it lowers blood pressure.”
Let denote the set of antecedents or conditions that must hold true in order for a law to be applicable to a particular data instance. Each condition ∈ is a prerequisite that must be satisfied for the corresponding law to be activated or relevant.
Example (in medical settings): • 1: “Patient has high blood pressure.” • 2: “Patient is over 60 years old.”
Let = {1, 2, . . . , } represent the set of input data fed into the AI system. Each ∈ represents a specific data sample.
Example (in medical settings): • 1: A data sample where the patient has high blood pressure and is 65 years old. • 2: A data sample where the patient has normal blood pressure and is 45 years old.
Let = {ℎ1, ℎ2, . . . , ℎ} represent the set of predictions or outcomes generated by the AI model. Each ℎ ∈ corresponds to a specific prediction for the instance .
Example (in medical settings): • ℎ1: “The patient is at high risk for cardiovascular disease.” • ℎ2: “The patient is not at high risk for cardiovascular disease.”
The key goal of the NDR framework is to use deductive reasoning to formalize how the AI model generates a prediction ℎ based on the combination of conditions and laws applied to the input . ∀ ∈ , ∃ℎ such that (1 ∧ 2 ∧ · · · ∧ ∧ 1 ∧ 2 ∧ · · · ∧ ) ⊢ ℎ Where: • ∈ is an input data instance. • 1, 2, . . . , are the conditions (e.g., patient characteristics like age, blood pressure). • 1, 2, . . . , are the domain laws (e.g., risk relationships).
• ⊢ denotes deductive reasoning leading to prediction ℎ.
Once we have the laws, conditions, and input data, the explanation for the prediction ℎ can be expressed as:
= (, , ) ⇒ ℎ • is the explanation for prediction ℎ. • is the function describing how , , and combine to produce ℎ.
• ⇒ indicates the logical flow from input to prediction.
In this research, the first step is to model the domain laws as continuous functions. These laws describe relationships between input variables and outcomes or predictions, and can be formalized as continuous functions that capture gradual changes in dynamic systems or continuous data. For example, in the medical domain: : R → R, () = some relationship between conditions 1() =
BloodPressure · Age 1 + Age ,
where = [BloodPressure, Age]
This equation models how the interaction between a patient’s blood pressure and age influences cardiovascular risk.
Conditions are modeled as constraints that must hold true for the corresponding laws to be applicable. These are represented as indicator functions: () =
0 otherwise {︃1 if condition holds true for instance Example (in medical settings): • 1() = {︃1 if blood pressure is high
0 otherwise
This ensures that the system only applies relevant laws when the conditions are satisfied.
Data instances are treated as vectors in an -dimensional space, representing diferent entities in the real world. For example, in healthcare, the vector could represent a particular patient’s characteristics. The data set is formalized as: Each represents an instance with features that are used in the model.
= (BloodPressure, Age, . . .)
∈ ⊂ R
The AI model generates a prediction ℎ as a function of the data instances, laws, and conditions. This can be represented as:
ℎ() = (1, 2, . . . , , 1, 2, . . . , , )
Where is a function that maps input data to a prediction. This function can take various forms, such as linear, non-linear, or other suitable formulations depending on the model’s complexity.
To formalize the inference process, we use calculus to express how the prediction ℎ() changes with respect to the input data . The derivative of ℎ() with respect to is computed as: ℎ() = ∑︁ ()
=1 · () applies only when the condition holds.
This derivative represents how sensitive the prediction is to changes in the input features. The term () represents how the laws change with respect to the input, and () ensures that the law
The explanation for a prediction ℎ can be generated by integrating over the laws and conditions that contributed to the prediction. This can be expressed as:
This integral sums over all possible laws and conditions, capturing the cumulative efect of how they interact with the data. In a probabilistic setting, we can also use a Bayesian approach: = ∫︁ ∫︁
(, , ) ∫︁
=
( | )
Where ( | ) represents the posterior probability of law given the data instance .
In AI systems, particularly those utilizing machine learning, the inference process is often optimized through a loss function. Hence, the optimization problem can be formalized as: =1 ℎˆ = arg min ∑︁
ℒ(ℎˆ(; ), )
Where ℒ is the loss function (e.g., mean squared error), ℎˆ is the predicted outcome based on model parameters , and is the true label. The parameters represent the weights associated with the laws, conditions, and other model components. We propose the architecture of the extended NDR framework as per the illustration in Figure 1.
Let’s consider the task of predicting the likelihood of a heart attack or stroke based on various risk factors. In this scenario, eNDR framework can be used to model causal relationships from cardiovascular theories, linking the disease to various risks. eNDR then applies the deductive reasoning on complex factors, handles uncertainty, and explains predictions based on the laws governing the medical domain (e.g., age, cholesterol, smoking habits). Any black-box learning model can be used to extract features from the dataset, and by eNDR the output is human-readable knowledge-based explanations.as captured on figure 2.
To validate the extended NDR framework with a focus on complex data and uncertainty, a complex health dataset (e.g., the Framingham Heart Study dataset) will be used. This dataset involves multiple features with both causal relationships and uncertainty (due to missing data, variable progression, and noise), which makes it an excellent candidate for demonstrating the efectiveness of eNDR in providing transparent and interpretable explanations grounded in causal knowledge.
The key metrics will include prediction accuracy to measure the algorithmic performance, uncertainty handling and rule coverage to measure its loyalty to Knowledge Base, as well as reasoning transparency and user trust measurement to evaluate eNDR trustworthiness.
Preliminary results show that integrating domain-specific laws into machine learning models enhances interpretability without compromising performance. Early tests in finance reveal that the NDR framework produces explanations aligned with human reasoning, ofering clear insights into decision-making. Unlike methods such as Causal Inference, Neuro-Symbolic Reasoning, Knowledge Graphs, LIME, and SHAP, which often cater to technical users, NDR focuses on intuitive, domain-grounded explanations, which is key to trust in AI-powered solutions. 6. Expected Next Research Steps and Final Contribution to Knowledge The next steps involve refining the NDR framework to handle larger, more complex datasets and incorporate probabilistic reasoning to account for uncertainty in real-world data. We will also evaluate the framework in additional domains, such as finance and criminal justice, to assess its generalizability.
The final contribution of this research will be a novel framework for Explainable AI that combines deductive reasoning with domain-specific knowledge, ofering a pathway for building more transparent and trustworthy AI systems. This work will also provide insights into how structured knowledge and causal reasoning can be embedded into machine learning models without compromising performance.
The author has used ChatGPT for grammar and content spelling check, after what he reviewed and edited the content as needed. The author takes full responsibility for the publication’s content [8] Miller, T. (2019). Explanation in Artificial Intelligence: Insights from the Social Sciences. Artificial
Intelligence, 267, 1-38.
[9] Preece, S., & Gunning, D. (2019). Explainable AI: A Survey of the State of the Art. IEEE Transactions
on Neural Networks and Learning Systems, 30(10), 2911-2924.
[10] Vilone, G.; Longo, L. Classification of Explainable Artificial Intelligence Methods through Their
Output Formats. Machine Learning and Knowledge Extraction 2021, 3(
Review. Journal of Healthcare Engineering, 2020, Article 8706534.
[14] Hempel, C.G.; Oppenheim, P. Studies in the Logic of Explanation. Philosophy of Science 1948, 15(
Artificial Intelligence. Artificial Intelligence Review , 53(