IHML for Oral Lesion Classification

• IHML integrates four calibrated base classifiers using a stacking meta-learning framework, explicitly mitigating intra-patient prediction variability.
• It fuses multimodal features—deep RGB, hyperspectral, texture, and demographic data—to enable robust four-class oral lesion classification.
• Patient-level posterior smoothing and uncertainty calibration via isotonic regression boost generalization, improving macro F1 by nearly 5%.

The incremental heuristic meta-learner (IHML) is a meta-learning framework introduced for robust four-class oral lesion classification using multimodal feature fusion and patient-aware uncertainty calibration. IHML operates as the final integrative stage of a larger oral lesion classification pipeline that incorporates deep RGB embeddings from ConvNeXt-v2, hyperspectral reconstructions, handmade spectral-textural descriptors, and demographic metadata. IHML is designed to fuse multiple heterogeneous base classifiers through calibrated probabilistic stacking and patient-level posterior smoothing, yielding a robust final prediction that explicitly mitigates intra-patient prediction variability and enhances generalization to unseen patients (Mukherjee et al., 15 Nov 2025).

1. Pipeline Integration and Objectives

IHML is deployed as the fourth and final stage in a multimodal lesion analysis pipeline. Stages 1–3 produce a multimodal feature vector $x \in \mathbb{R}^D$ for each lesion ROI, encompassing deep RGB features, hyperspectral biomarkers, texture, spectral-shape, and clinical attributes. In stage 4, IHML leverages $M=4$ calibrated base classifiers—LightGBM, Extra Trees, Gradient Boosting, and Logistic Regression—each trained on these feature vectors. The outputs of these base classifiers are then integrated via the IHML strategy to generate the final prediction $\hat{y}$.

The principal objectives of IHML are:

1. To fuse heterogeneous base models in an uncertainty-aware and calibrated manner.
2. To reduce patient-level prediction variability through probabilistic smoothing.
3. To yield reliable four-class decisions (healthy, benign, OPMD, OCA) that generalize robustly to previously unseen patients (Mukherjee et al., 15 Nov 2025).

2. Algorithmic Structure and Mathematical Formulation

At its core, IHML operates as a stacking meta-learner with explicit uncertainty modeling and group-level regularization. The methodology comprises the following key algorithmic stages:

Calibration and Confidence Feature Extraction

Each base classifier $g_m$ provides raw score vectors $s_i^{(m)} \in \mathbb{R}^4$ for sample $i$. These are transformed through a calibration mapping $\text{Cal}_m: \mathbb{R}^4 \to \Delta^4$ (using isotonic regression) to obtain probability vectors $p_i^{(m)} = \text{Cal}_m(s_i^{(m)})$.

For each $p_i^{(m)}$, a confidence feature vector $c_i^{(m)}$ is constructed as:

$$\psi(p_i^{(m)}) = \begin{bmatrix} \max_{c} p_{i,c}^{(m)} \ p_{i,c_1}^{(m)} - p_{i,c_2}^{(m)} \quad (c_1, c_2 = \text{top 2 classes}) \ -\sum_{c=1}^4 p_{i,c}^{(m)} \log p_{i,c}^{(m)} \end{bmatrix} \in \mathbb{R}^3$$

Patient-Level Posterior Smoothing

Probabilities for samples belonging to the same patient (indexed by $g(i)$) are iteratively smoothed:

$$p_i^{(t+1)} = (1-\alpha)p_i^{(t)} + \alpha\,\overline{p_{g(i)}^{(t)}}$$

where $\overline{p_{g(i)}^{(t)}}$ is the mean probability across all samples from patient $g(i)$. This process is repeated for $T$ iterations to mitigate prediction inconsistency within patient groups.

Meta-Classifier Training

Features for meta-classification are formed as:

$$h_i = [p_i^{*(1)};\dotsc;p_i^{*(M)};c_i^{(1)};\dotsc;c_i^{(M)}] \in \mathbb{R}^{M \cdot (K+4)}$$

where $p_i^{*(m)}$ are post-smoothing probabilities. A multinomial logistic regression with softmax normalization outputs final class probabilities:

$$q_i = \text{softmax}(Wh_i + b), \qquad \hat{y}_i = \arg\max_{c} q_{i,c}$$

3. Training, Inference, and Incremental Update

Training Protocol

• Base models are trained on 85% of patient data using 5-fold patient-wise cross-validation.
• Calibration is performed on held-out CV splits, ensuring proper probability mapping and avoiding overfitting.
• Smoothing hyperparameters ($\alpha$, $T$) are selected based on validation.
• The final meta-classifier is trained on concatenated (smoothed) probabilities and confidence features.

Inference and Incremental Use

At inference, each test sample:

1. Receives calibrated probability predictions and confidence features from each base model.
2. Undergoes $T$ rounds of patient-level smoothing.
3. Feeds concatenated, smoothed probabilities and static confidences into the meta-classifier for final prediction.

Incremental update is supported: New patient images can be directly incorporated by recomputing group means and reapplying the smoothing and meta-classification steps, without full retraining of base learners or the meta-classifier.

Step	Input	Output
Calibration	$s_i^{(m)}$	$p_i^{(m)}$ (calibrated probability)
Confidence Statistics	$p_i^{(m)}$	$c_i^{(m)}$ (confidence vector)
Posterior Smoothing	$p_i^{(m)}$	$p_i^{*(m)}$ (smoothed probability)
Meta-Classification	$h_i$	$q_i$, $\hat{y}_i$ (final class & probabilities)

4. Hyperparameters and Design Considerations

Key hyperparameters and their impact are as follows:

• Number of base learners $M=4$: LightGBM, Extra Trees, Gradient Boosting, and Logistic Regression.
• Calibration: Isotonic regression provides monotonic probability mapping for all base models.
• Confidence features per base $K=3$: Maximum probability, top-two margin, and Shannon entropy.
• Smoothing parameters: $\alpha \in \{0.1, 0.3, 0.5\}$, with $\alpha=0.3$ providing an optimal stability-accuracy balance; $T \in \{1,\dots,5\}$, with $T=3$ balancing convergence and runtime.
• Meta-classifier regularization: L2 penalty $C=1.0$ tuned via validation.

Larger values of $\alpha$ increase consistency within patient predictions but risk over-smoothing genuine inter-lesion differences. More smoothing iterations $T$ exploit group priors up to a point, then yield diminishing gains. Calibration and confidence statistics each contributed approximately 1–2% absolute macro F1 improvement in ablation experiments (Mukherjee et al., 15 Nov 2025).

5. Empirical Performance and Ablation Analysis

Ablation studies and comparative analyses on unseen patient data demonstrate the efficacy of IHML. The best individual base model (LightGBM) achieved a macro F1 of 61.27%. Applying the full IHML pipeline increased macro F1 to 66.23% (a +4.96% improvement), with accuracy rising from 59.81% to 64.56% and AUC-ROC from 82.89% to 84.45% (+1.56%).

Feature-group ablation within IHML showed progressive gains:

• ConvNeXt only: F1 = 54.97%
• +demographics: F1 = 60.08%
• +haemoglobin-sensitive features: F1 = 63.19%
• +texture: F1 = 64.13%
• +spectral-shape (full IHML): F1 = 66.23%

These results confirm that both calibrated stacking and patient-level smoothing are critical for achieving robust uncertainty-aware classification. The framework also demonstrates substantial improvement over naive feature fusion and isolated learners (Mukherjee et al., 15 Nov 2025).

6. Context and Significance

IHML is specifically targeted at oral lesion classification in real-world, low-resource settings, where annotated multimodal clinical data are limited and intra-patient variations can confound predictions. By integrating calibrated probabilistic stacking with group-level posterior smoothing, IHML enforces patient-level coherence in multi-image diagnostic inference. This design is particularly suited to applications requiring uncertainty-aware decision making and efficient incorporation of new patient data without full retraining.

The approach demonstrates the importance of uncertainty calibration, explicit group-level regularization, and lightweight meta-classifiers in heterogeneous data fusion pipelines. A plausible implication is that similar strategies could be adapted to other structured medical imaging or multi-modal scenarios where group-level context is available and critical for robustness (Mukherjee et al., 15 Nov 2025).