The ability to interpret machine learning model decisions is critical in such domains as healthcare, where trust in model predictions is as important as their accuracy. Inspired by the development of prototype parts-based Deep Neural Networks in computer vision, we propose a new model for tabular data, specifically tailored to medical records, that requires discretization of diagnostic result norms. Unlike the original vision models that rely on the spatial structure, our method employs trainable patching over features describing a patient, to learn meaningful prototypical parts from structured data. These parts are represented as binary or discretized feature subsets. This allows the model to express prototypes in human-readable terms, enabling alignment with clinical language and case-based reasoning. Our proposed neural network is inherently interpretable and ofers interpretable concept-based predictions by comparing the patient's description to learned prototypes in the latent space of the network. In experiments, we demonstrate that the model achieves classification performance competitive to widely used baseline models on medical benchmark datasets, while also ofering transparency, bridging the gap between predictive performance and interpretability in clinical decision support.
Machine learning (ML) has been increasingly used in medicine for many decades, in particular to improve
diagnostic accuracy, predict patient outcomes, and support clinical decision making by uncovering
complex patterns in medical data [
Nowadays, in many tasks, Deep Neural Networks (DNN) have become the most popular approach,
particularly for analyzing modalities such as images, time series, or text data. However, a significant
portion of clinical work still relies on tabular data, where the application of deep learning models,
due to their black-box nature, is less widespread and less appreciated. In the healthcare domain, the
reluctance to adopt DNN is partially driven by the dificulty in interpreting their decision-making
processes, making it challenging for physicians to analyze, validate, and ultimately trust their results
in real-world applications. As a result, there has been growing interest in using Explainable AI (XAI)
techniques [
Currently, the landscape of XAI is dominated by feature importance methods such as SHAP [
In this context, prototypes – instances that represent groups of similar examples – have emerged as a
particularly promising explanation technique [
Inspired by the paper [
To address this gap, we propose a prototype-based neural network, called Model for Explainable Diagnosis using Interpretable Concepts (MEDIC), specifically designed for tabular medical data. Our approach introduces discrete prototypes, with the aim of improving interpretability while maintaining strong predictive performance. While traditional models such as DTs ofer symbolic interpretability, their reliance on rigid rule structures may not align well with the complexity of clinical reasoning. In contrast, our method adopts a prototype-based approach that enables more flexible, example-driven explanations, allowing clinicians to interpret decisions through similarities to real, representative patient cases. The goal of this study is to develop and evaluate this model in the context of medical records of patients, with a focus on producing faithful and physician-friendly explanations.
To ensure reproducibility, the code is publicly available in a GitHub repository1.
The work [
Despite machine learning models consistently demonstrating superior performance with tabular
clinical data, these ensemble methods sufer from inherent opacity in their decision processes, creating
a critical need for efective explanation frameworks that can provide healthcare professionals with
transparent insights into model reasoning [
The landscape of explainable AI approaches can be broadly categorized into two paradigms: feature
attribution methods and concept-based methods [
Prototype-based explanations generally fall into two families. The post-hoc family identifies
prototypes after model training, typically selecting representative instances from the training set. Notable
algorithms include MMD-Critic [
datasets containing many irrelevant features, which is a common characteristic in healthcare, where
comprehensive diagnostic panels frequently generate information records containing redundant
information [
The second family, ante-hoc or intrinsic prototype methods, integrates prototype reasoning directly
into the model architecture. Usually, these approaches represent prototypes not as complete instances
but as parts or feature conjunctions that participate in decision making through mechanisms such as
weighted voting. This direction gained significant attention following [
ProtoPNet [
However, despite ProtoPNet’s successful application across various image processing tasks [
In this section, for the first time, we propose the neural network MEDIC: Model for Explainable Diagnosis
using Interpretable Concepts. MEDIC is inspired by the prototypical parts paradigm proposed in [
At a high level, the architecture follows an interpretable processing pipeline consisting of four key stages: (1) input discretization, which transforms continuous variables in the patient’s description into symbolic bins; (2) part extraction, which identifies sparse and semantically coherent subsets of input features; (3) prototype comparison, where each extracted part of the patient’s description is matched to learned prototypes, stored as embeddings representing features subsets of feature-value pairs from the training data; and (4) classification of the considered instance based on its similarity to prototypes. The complete MEDIC model is trained end-to-end to jointly learn all of these components in a supervised setting.
In clinical data, where features often come from heterogeneous sources (e.g. lab tests values, vital signs, diagnoses), such structured reasoning aligns well with domain expert expectations. Discretized bins can reflect clinically relevant ranges of diagnostic tests (e.g., abnormally high glucose), sparse parts mirror combinations that physicians would consider jointly (e.g., elevated CRP and fever), and prototypes anchor predictions in real cases that can be inspected post hoc.
We now describe the architecture in detail, starting with describing the interpretable discretization of continuous input features.
(a) Fuzzy binning: the input value is softly as- (b) Hard binning: the input is deterministically signed to each bin based on proximity to bin’s assigned to a single bin with the nearest center. center. The final encoding is a weighted combi- The encoding becomes a one-hot vector. nation, where the weights reflect similarity to the bin centers.
The decretization of continuous medical variables (e.g. age, lab tests values) into symbolic categories can aid interpretation and facilitate reasoning about patient features. In this work, our aim is to ultimately produce ranges of continuous features for symbolic interpretability. However, such a hard discretization is hardly optimizable in gradient-based neural network training.
To overcome this challenge, we introduce a fuzzy binning layer that enables a smooth, diferentiable approximation of hard discretization during training. This allows gradients to flow through the discretization process and enables end-to-end optimization. After training, the soft representation can be replaced with a deterministic hard binning for better interpretability.
Fuzzy Binning To allow interpretable discretization of continuous input features, we introduce a fuzzy binning layer that softly assigns each scalar feature value ∈ R to a set of trainable bins. Each bin is characterized by a learnable center ∈ R and a shared bandwidth parameter > 0 . The soft membership of to the bin is defined using a Gaussian kernel: () = ( − 2 2
)2 ˜() =
exp(− ()) ∑︀=1 exp(− ()) + (1) (2) where ˜() denotes the normalized soft assignment and is a small constant added for numerical stability. This results in a fuzzy, probabilistically weighted representation over bins, allowing each input to contribute partially to multiple bins (see Figure 1a).
The use of Gaussian kernels for fuzzy binning ofers several advantages over direct distance-based assignment (e.g., 2 norm). First, the smooth exponential decay naturally reflects uncertainty in proximity, which is especially relevant when feature values lie near bin boundaries. Second, the resulting softmax distribution is diferentiable and normalized, facilitating gradient-based optimization in Deep Neural Networks.
Importantly, the bin centers and shared bandwidth are optimized jointly with other model parameters during end-to-end training, allowing the discretization scheme to adapt to the data distribution.
After initial training of the network, the discretization is switched to the hard mode. In the hard setup, the input is assigned to a single bin via a non-diferentiable arg min operation over squared distances: (3) (4) ︂( ^() = one_hot arg min ( − ) 2 ︂)
The resulting representation is a one-hot2 vector (Figure 1b), which can be advantageous for symbolic interpretation and comparison of prototypes. However, it lacks gradient flow, making it unsuitable for end-to-end training.
In the hard binning regime, the input feature values are partitioned into contiguous intervals derived from the learned bin centers { }=1. Specifically, each bin is associated with the interval = ⎧(︀ ⎪ ⎨︀[ −1 ⎪⎩[︀ −1 − ∞, 1+ 2 )︀
2 2 + , + +1 )︀
2 2 + , +∞)︀ if = 1 if 1 < < if = facilitates interpretation by decision makers. such that any scalar input is discretized into the bin for which ∈ . This alternative representation
MEDIC is a neural network inspired by the interpretable prototypical parts-based classification paradigm
proposed in [
In the beginning, the raw input vector of features describing the patient is transformed by the network into a sparse high-dimensional binary representation. Each continuous feature is processed by a binning module introduced in Section 3.1. Meanwhile, categorical features undergo one-hot2 encoding. All vectors coming from discretization are concatenated into a single ′ dimensional vector.
Next, the binarized input is multiplied with a trainable set of patching masks, encoded as a matrix ∈ R× ′ . Each mask selects and linearly combines a sparse subset of binary features, efectively defining a part of the input instance (i.e. its description by features). Intuitively, each part can be seen as a meaningful combination of clinical indicators – for example, high blood pressure in elderly patients or elevated glucose and BMI. This encourages the model to focus on patterns that are not only interpretable but also structured in a way that reflects domain knowledge.
Each of the part vectors is passed through a shared feature extractor module, implemented as a shallow multilayer perceptron with ReLU activations. This module transforms each sparse binary part into a dense embedding – a compact vector of size ℎ that captures abstract and informative features. Embeddings are designed to preserve meaningful relationships in the data while reducing dimensionality, enabling the model to generalize across similar patterns. From an interpretability perspective, this step summarizes each clinically relevant pattern into a low-dimensional representation that retains the most diagnostically informative aspects.
To facilitate interpretable decision-making, the network maintains a set of learnable prototype vectors of size ℎ. For each input part represented as embedding, the 2 distance is computed for every prototype. This results in a ×
distance matrix, where each entry quantifies the dissimilarity between a specific part and a prototype. A max-pooling operation across parts selects the most relevant part for each prototype, yielding a vector of minimal distances, where each entry reflects the smallest distance between a given prototype and the most similar embedding representing a part of the input 2A one-hot vector is a way of representing categories or intervals where only one entry is "on" (set to 1) and all others are will be marked as active. This makes it easy to interpret into which clinical range the value falls. describing the patient. This enables comparison of each prototype to its best-matching part in the patient description.
To enable case-based reasoning, the network maintains a set of learnable prototype vectors, each of dimension ℎ. Conceptually, each prototype represents a summary of a typical clinical condition or patient case learned from data. Each prototype is anchored in real patient data and corresponds to a representative example that lies near the center of a cluster of similar cases, making it reflective of common patterns observed across many patients. For every embedding corresponding to the part of patient description, the model computes the squared Euclidean (2) distance to each prototype, yielding a × distance matrix. Each row corresponds to one patient description part and each column to a prototype.
Then a maximum-pooling operation is applied across the rows of this matrix (that is, across parts), selecting for each prototype, the input part that has the smallest distance to that prototype, efectively identifying the most similar part. This produces a distance vector of length , which summarizes how closely the input aligns with each of the learned prototypes.
Finally, this vector of distances is passed through a linear classification layer, producing a probability distribution over target classes. Since the classification decision is based directly on similarity to interpretable prototypes, each linked to specific input parts, the resulting predictions can be traced back and explained in terms of clinically meaningful comparisons to learned prototype parts.
To ensure stable and interpretable network training, we adopt a three-stage training procedure that enables the model to learn hard bins – intervals, and realistic prototype parts directly from the training data, all within a gradient-based optimization framework.
Stage 1: Initialization with Fuzzy Binning and Learnable Prototypes In the initial stage, the entire network is trained end-to-end with fuzzy binning and randomly initialized learnable prototypes. This setting ensures smooth gradient flow through the discretization modules, allowing the network to co-adapt binning thresholds and part extraction masks.
Fuzzy binning uses soft Gaussian kernels (Section 3.1, Figure 1a), which provide fuzzy assignments across bins. The patching masks and prototypes are trained jointly using classification loss, combined with auxiliary regularization terms: (1) L1 sparsity of patching masks, and (2) a diversity penalty to encourage spread among prototypes, which are further discussed in Section 3.4.
Stage 2: Hard Binning and Mask Discretization Once convergence in the training criterion is achieved, the discretization mode is switched to a hard mode by replacing the fuzzy binning with hard arg min bin selection, freezing binning thresholds (Section 3.1, Figure 1b). Additionally, patching masks are binarized by thresholding to enforce strict binary groupings of input dimensions into parts.
This transition enables symbolic interpretability and highlights which specific input features are most relevant for each part. The rest of the network is fine-tuned using discretized inputs, preserving the interpretability of the parts.
Stage 3: Prototype Replacement with Real Parts Finally, the learned prototypes are replaced with embeddings derived from parts of actual patient records in the training data, ensuring that each prototype corresponds to a real and representative clinical case. For each prototype, the closest embedded input part is identified using the 2 distance. These real parts are then copied into the prototype memory, replacing the synthetic prototypes. This step improves interpretability by anchoring each prototype to an actual example from the data.
This last step grounds the network’s reasoning in actual data, allowing domain experts to inspect prototypical cases for each class. During this phase, the prototype embeddings are frozen, and only the classification head is fine-tuned to maintain stable performance, as accuracy would otherwise be expected to decline.
Model and training complexity From a computational standpoint, MEDIC maintains a relatively simple neural network architecture, comparable in training complexity to a shallow multi-layer perceptron with five layers. The operations within the network consist primarily of matrix multiplications and element-wise products, which are fully parallelizable on modern hardware and avoid sequential dependencies inherent in architectures such as Recurrent Neural Networks. The final stage of prototype selection, which matches learned embeddings to parts of real patient records, scales linearly with the size of the dataset, making it eficient even for larger collections. As a result, both training and inference remain computationally lightweight, with performance characteristics similar to other small feed-forward neural networks.
The model is trained using cross-entropy loss, later denoted as ℒ , as the standard objective for classification tasks. To improve interpretability and promote eficient structure, two regularization terms are added. The first is an ℓ 1 sparsity penalty applied to the patching mask parameters: ′ 1 ∑︁ ∑︁ ℒsparsity = sparsity · ℓ 1( ) = sparsity · ′ =1 =1 | | , (5) where ∈ R× ′ are the patching masks. This term encourages parts to rely on a minimal set of input features describing each patient. The ℓ1 penalty is chosen because, unlike ℓ2, it promotes exact zeros in patching masks, efectively turning of irrelevant input dimensions, and therefore leading to sparser and therefore more interpretable part-feature associations.
The second term encourages diversity among prototypes by penalizing redundancy in their representations: 1 diversity · ( − 1) ℒdiversity = − ∑︁ ‖z − z ‖2 , (6) ̸= where z and z are prototype embeddings. This promotes coverage of distinct regions in the latent space. The full training objective function is the sum of three above-mentioned loss parts: ℒ = ℒCE + ℒsparsity + ℒdiversity . (7)
This section presents a comprehensive evaluation of our model, both from an interpretability and a predictive performance perspective. In the beginning, in Section 4.1 we assess the predictive accuracy of the method on three benchmark datasets, comparing its performance to selected baseline models. Subsequently, in Section 4.2 we demonstrate how the learned prototypes may be applied in practice, through a case study grounded in a real-world medical dataset. This analysis highlights the interpretability of MEDIC and its ability to form clinically plausible representations.
4.1.1. Experimental setup Data To evaluate the proposed model, three publicly available medical datasets were selected: Cirrhosis3, Chronic Kidney Disease (CKD)4, and Diabetes5. These datasets were chosen due to their clinical relevance and inclusion of multiple laboratory measurements such as blood test results, namely: • Cirrhosis: bilirubin, cholesterol, albumin, copper, triglycerides, alkaline phosphatase (ALP), serum glutamic-oxaloacetic transaminase (SGOT), platelets, and prothrombin time; • CKD: red blood cells, pus cells, pus cell clumps, blood glucose, blood urea, serum creatinine, sodium, potassium, hemoglobin, packed cell volume, white blood cell count, and red blood cell count; • Diabetes: blood glucose and insulin.
Enumerated tests are well suited for discretization. These datasets also include additional numerical
indicators, such as body mass index (BMI, in the Diabetes dataset), which further benefit from
discretization by enhancing interpretability. Moreover, all three datasets exhibit a class imbalance: Cirrhosis
contains 125, 19, and 168 instances for classes death, censored, and censored due to liver transplantation
respectively; CKD dataset consists of 115 and 43 instances for classes not CKD and CKD; and Diabetes
includes 500 negative and 268 positive samples for diabetes presence. These characteristics present
a realistic benchmark for evaluating the model’s ability to process numerical medical features while
addressing class imbalance, a common challenge in clinical predictive modeling [
Criterion To compare the performance of diferent models, we use the geometric mean (g-mean) of sensitivity and specificity, defined as: 3https://archive.ics.uci.edu/dataset/878/cirrhosis+patient+survival+prediction+dataset-1 4https://archive.ics.uci.edu/dataset/336/chronic+kidney+disease 5https://www.kaggle.com/datasets/mathchi/diabetes-data-set 6https://scikit-learn.org/ 7https://xgboost.readthedocs.io/ g-mean = √︀sensitivity × specificity (8) where sensitivity (also called recall) measures the proportion of actual positive cases correctly identified and specificity measures the proportion of actual negative cases correctly identified: sensitivity = , specificity = (9)
+ +
The g-mean balances performance on both classes, ensuring the model is not biased toward the
majority class. This is especially useful in medical datasets with class imbalance, where one outcome
(e.g., disease presence) is much rarer. Unlike accuracy, g-mean provides a more balanced and clinically
meaningful measure by ensuring good performance on both positive and negative classes [
For MEDIC, we recommend a structured tuning strategy informed by prior experience. Begin by
setting sparsity to a relatively high value (e.g. ≈ 1.0 ) and diversity = 0, together with a large number of
prototypes (e.g. 64). Gradually decrease sparsity until the number of activated features in the prototype
parts stabilizes within a comprehensible range, ideally fewer than 5-7 features per prototype part.
Once this is achieved, incrementally increase diversity to promote diversity inactivated parts, ensuring
that the prototype part lengths remain consistent. After arriving at interpretable and stable prototype
configurations, the remaining hyperparameters, including the number of prototypes (see Table 1) can
be automatically tuned using a hyperparameter optimization algorithm such as TPE [
Table 1 summarizes the hyperparameters tuned for each model and the corresponding search spaces. For implementation-specific details, we refer the reader to the respective baseline packages cited in the Baselines paragraph above. 4.1.2. Results The performance of the model is summarized in Table 2, using the geometric mean (g-mean) metric in three datasets. MEDIC demonstrates competitive performance, achieving the best g-mean on the Cirrhosis and CKD datasets. In the diabetes data set, although XGB achieved the highest score, MEDIC followed closely, within less than a percentage point, indicating comparable efectiveness.
Table 3 shows the maximum allowed number of prototypes defined as a model setting (hyperparameter) and the number of unique prototype parts actually discovered by the MEDIC model during training. The results suggest that the model can self-regularize by reusing the same prototype multiple times when no additional meaningful feature-value sets (prototype parts) can be identified. This indicates that MEDIC avoids overfitting by focusing only on truly informative patterns, even when more prototypes are allowed.
To show that MEDIC is interpretable and present how prototype parts learned by MEDIC look like, we conducted a qualitative analysis using a case study for a single dataset – Cirrhosis. Our aim is to illustrate how the MEDIC reasoning process works by comparing individual patient cases to representative clinical patterns (prototypes) previously learned from the data.
Table 4 shows the discretized feature intervals identified by the network. These intervals represent meaningful partitions of the input space and often align with known clinical thresholds. For reference, we compare them with the standard clinical intervals provided by the American College of Clinical Pharmacy9.
For example, the learned limit between intervals 1 and 2 for albumin is 3.7 g/dL, which closely matches the clinical lower limit of 3.5 g/dL. Similarly, the learned limits for the prothrombin time (10.52-10.93 seconds) are well within the reference range of 10–13 seconds. Triglycerides also have a limit near 137 mg / dL, close to the reference value of <150 mg/dL.
Earlier in this work, we justified using three bins per feature to intuitively capture low–normal–high
ranges. However, for certain features in specific disease contexts, deviations in one single direction may be clinically significant. Interestingly, in this experiment the observations suggest the network to exhibit the ability to self-organize and adjust these bins accordingly. For example, for copper, the first interval is efectively disabled by learning a negative upper bound (–8.98), which is not physiologically plausible, thus disregarding it. Likewise, for platelets and albumin, the network forms exceptionally narrow middle intervals, suggesting that small changes within this range may be critical for the classification.
Although Table 4 shows intervals ranging from −∞ to ∞ for technical completeness, in practical applications these can be translated into clinically relevant and bounded intervals. For example, lower limits can be set to zero and upper limits can be capped according to known physiological limits, without afecting the model’s learning performance. This translation can support physicians in interpreting the model’s behavior more easily.
Subsequently, MEDIC identified several prototype parts, specific combinations of clinical features and value ranges, that it considers informative to predict patient outcomes related to cirrhosis. These prototype parts are presented in Table 5.
Next, we investigate how a specific patient case (shown in Table 6, classified as 0 – death) is internally processed and classified by MEDIC through its similarity to the nearest learned prototype parts. Each prototype part consists of a sparse conjunction of conditions over discretized or binary features, typically involving only a small number of dimensions. For example, a prototype part may specify conditions such as: Bilirubin level within [0.79, 3.43) mg/dL, absence of hepatomegaly (Hepatomegaly = 0), and drug usage indicated (Drug = 1). These concise feature subsets capture clinically meaningful patterns that contribute to the model’s decisions. MEDIC’s classification is driven by the similarity between the patient’s description and the prototype parts, which are easily accessible and can be examined by the user, thereby ofering transparent insight into the model’s reasoning process.
The list of prototypes with the highest similarity to this example demonstrates how the model constructs its reasoning by combining interpretable substructures. Many of these substructures align with known clinical heuristics or highlight relevant feature interactions. For example, bilirubin, ALP (alkaline phosphatase), and N_Days (duration since patient registration) appear frequently in the most similar prototypes, highlighting their importance as clinical indicators influencing classification. Bilirubin in [0.79, 3.43) ∧ Hepatomegaly = 0 ∧ Spiders = 0 Albumin ∈ [3.82, ∞) ∧ Bilirubin ∈ [0.79, 3.43) ∧ Hepatomegaly = 0 ∧ Spiders = 0 Cholesterol ∈ [667, ∞) ∧ Copper ∈ [103.76, ∞) ∧ Hepatomegaly = 0 ∧ Spiders = 0 Bilirubin ∈ (−∞ , 0.79) ∧ Cholesterol ∈ (−∞ , 345) ∧ Hepatomegaly = 1 ∧ Spiders = 0 Albumin ∈ [3.70, 3.82) ∧ Bilirubin ∈ [0.79, 3.43) ∧ Cholesterol ∈ (−∞ , 345) ∧ Hepatomegaly = 0 ∧ Spiders = 0 Bilirubin ∈ [0.79, 3.43) ∧ Cholesterol ∈ (−∞ , 345) ∧ Hepatomegaly = 0 ∧ Platelets ∈ (−∞, 271) ∧ Spiders = 0 Bilirubin ∈ (−∞ , 0.79) ∧ Cholesterol ∈ (−∞ , 345) ∧ Hepatomegaly = 0 ∧ Platelets ∈ (−∞, 271) ∧ Spiders = 0 Albumin ∈ [3.70, 3.82) ∧ Bilirubin ∈ [0.79, 3.43) ∧ Cholesterol ∈ (−∞ , 345) ∧ Hepatomegaly = 0 ∧ Spiders = 0 Albumin ∈ [3.82, ∞) ∧ Bilirubin ∈ [0.79, 3.43) ∧ Cholesterol ∈ (−∞ , 345) ∧ Hepatomegaly = 1 ∧ Spiders = 0 ALP ∈ (−∞ , 3668) ∧ Bilirubin ∈ [0.79, 3.43) ∧ Hepatomegaly = 0 ∧ N_Days ∈ [2343, ∞) ∧ SGOT ∈ [80, 144) ∧ Tryglicerides ∈ (−∞, 137) ALP ∈ (−∞ , 366) ∧ Bilirubin ∈ [0.79, 3.43) ∧ Drug = 1 ∧ Hepatomegaly = 0 ∧ N_Days ∈ (−∞, 2152) ∧ SGOT ∈ [80, 144) ∧ Tryglicerides ∈ [172, ∞) Albumin ∈ [3.82, ∞) ∧ ALP ∈ (−∞ , 3668) ∧ Drug = 1 ∧ Hepatomegaly = 0 ∧ N_Days ∈ [2343, ∞) ∧ SGOT ∈ (−∞, 80) ∧ Tryglicerides ∈ [172, ∞) ALP ∈ (−∞ , 3668) ∧ Bilirubin ∈ [0.79, 3.43) ∧ Drug = 1 ∧ Hepatomegaly = 0 ∧ N_Days ∈ (−∞ , 2152) ∧ Prothrombin ∈ (−∞ , 10.52) ∧ SGOT ∈ [80, 144) ∧ Tryglicerides ∈ [172, ∞) Albumin ∈ [3.82, ∞) ∧ ALP ∈ (−∞ , 3668) ∧ Bilirubin ∈ [0.79, 3.43) ∧ Drug = 1 ∧ Hepatomegaly = 0 ∧ N_Days ∈ [2343, ∞) ∧ Prothrombin ∈ (−∞ , 10.52) ∧ SGOT ∈ [80, 144) ∧ Tryglicerides ∈ (−∞, 137)
These prototypes highlight clinically relevant signals, such as low bilirubin, no hepatomegaly, and shorter hospital stays (N_Days), as contributors to the classification. Furthermore, the inclusion of interaction patterns, such as elevated triglycerides in the context of certain ranges of liver enzymes, reflects how the network captures more nuanced decision logic than simple thresholding. 1. Similarity: 0.864
Prototype: N_Days ∈ (−∞, 2152) ∧ Drug = 1 ∧ Hepatomegaly = 0 ∧ Bilirubin ∈ [0.79, 3.43) ∧ ALP ∈ (−∞, 366) ∧ SGOT ∈ [80, 144) ∧ Tryglicerides ∈ [172, ∞) 2. Similarity: 0.846
Prototype: Hepatomegaly = 0 ∧ Spiders = 0 ∧ Cholesterol ∈ [667, ∞) ∧ Copper ∈ [103.76, ∞) 3. Similarity: 0.834
Prototype: N_Days ∈ [2343, ∞) ∧ Hepatomegaly = 0 ∧ Bilirubin ∈ [0.79, 3.43) ∧ ALP ∈ (−∞, 3668) ∧ SGOT ∈ [80, 144) ∧ Tryglicerides ∈ (−∞, 137) 4. Similarity: 0.824
Prototype: N_Days ∈ [2343, ∞) ∧ Drug = 1 ∧ Hepatomegaly = 0 ∧ Bilirubin ∈ [0.79, 3.43) ∧ Albumin ∈ [3.82, ∞) ∧ ALP ∈ (−∞, 3668) ∧ SGOT ∈ [80, 144) ∧ Tryglicerides ∈ (−∞, 137) ∧ Prothrombin ∈ (−∞, 10.52) 5. Similarity: 0.798
Prototype: N_Days ∈ (−∞, 2152) ∧ Drug = 1 ∧ Hepatomegaly = 0 ∧ Bilirubin ∈ [0.79, 3.43) ∧ ALP ∈ (−∞, 3668) ∧ SGOT ∈ 80, 144) ∧ Tryglicerides ∈ [172, ∞) ∧ Prothrombin ∈ (−∞ , 10.52) Although MEDIC’s prototype parts may resemble rules derived from DTs, they difer fundamentally in how they are used for decision making. DTs require strict rule satisfaction, whereas MEDIC allows partial matches to prototype parts, enabling more flexible and probabilistic reasoning. This tolerance to incomplete matches can improve robustness to noise, missing values, and borderline cases, which are common issues in clinical data due to measurement variability, incomplete testing, or inconsistent documentation.
This work introduced MEDIC, a novel prototype parts-based neural network architecture that transforms the approach to interpretability in machine learning for medical tabular data. Unlike conventional post-hoc explanation methods that retrospectively justify black-box decisions, MEDIC represents a paradigm shift toward inherently interpretable models that mimic clinical reasoning patterns. The core innovation lies in our three-component architecture: (1) diferentiable discretization that aligns with medical thresholds, (2) sparse patching masks that identify clinically meaningful feature combinations, and (3) prototype-based reasoning that grounds predictions in case-based comparisons, all unified within an end-to-end trainable framework.
Evaluation across three clinical datasets demonstrated that MEDIC achieves competitive and sometimes superior predictive performance compared to established methods while providing transparent decision processes. In particular, the model autonomously discovered discretization thresholds that closely align with clinically recognized reference ranges, as evidenced in our cirrhosis case study where albumin and prothrombin intervals closely matched established medical guidelines. Furthermore, the prototype parts learned by the model reflected combinations of features that correspond to recognizable diagnostic patterns, suggesting that MEDIC captures meaningful representations of clinical knowledge.
The implications of this work extend beyond technical innovation only. By bridging the gap between accuracy and interpretability, MEDIC addresses a critical barrier to AI adoption in healthcare, the lack of interpretability, which undermines the trust of clinicians and regulatory acceptance. Our approach supports collaborative human-AI decision making where the model’s reasoning can be verified, critiqued, and integrated with clinical expertise.
The interpretability of MEDIC is achieved by grounding each prediction in prototypical parts – concise, clinically meaningful feature patterns drawn from real patient data and presented in natural, domain-specific language. Such clarity is essential for building trust, enabling clinicians to understand and validate the model’s reasoning, and ensuring that AI-assisted decisions can be confidently integrated into medical practice.
Several promising directions emerge for future research. First, incorporating domain-specific prior
knowledge into the prototype learning process could further align the model’s representations with
established medical understanding. Second, investigating methods for dynamic prototype adaptation
could enable the model to update its learned representations in response to changes in symptoms over
time. This drift in symptoms may result from evolving disease variants, treatment efects, or changes in
how diseases present between populations, as seen with COVID-19 [
This research was funded in part by National Science Centre, Poland OPUS grant no. 2023/51/B/ST6/00545 and in part by PUT SBAD 0311/SBAD/0752 grant.