SwasthLLM: Multilingual Medical Diagnosis

• SwasthLLM is a multilingual medical diagnosis framework that applies zero-shot, cross-lingual, and multi-task learning to classify 24 diseases from clinical symptom narratives in English, Hindi, and Bengali.
• It integrates XLM-RoBERTa with a language-aware attention mechanism, Siamese contrastive learning, and MAML to align representations across languages and boost diagnostic accuracy.
• Empirical evaluations show 97% test accuracy in English and robust zero-shot performance in Hindi and Bengali, highlighting its potential for low-resource clinical settings.

Searching arXiv for the named paper and closely related work to ground the article in current literature. SwasthLLM is a multilingual medical-text diagnosis framework introduced as a unified, zero-shot, cross-lingual, and multi-task learning framework for automatic disease diagnosis from clinical text across English, Hindi, and Bengali without language-specific fine-tuning. It is motivated by the scarcity of annotated medical data in low-resource languages, linguistic variability across populations, and the limitations of translation-dependent pipelines in healthcare. At its core, SwasthLLM combines XLM-RoBERTa-base, a language-aware attention mechanism, a disease classification head, a Siamese contrastive learning module, a translation consistency module, and Model-Agnostic Meta-Learning (MAML) to align medically relevant representations across languages while preserving diagnostic performance (Sar et al., 24 Sep 2025).

1. Definition and problem setting

SwasthLLM addresses automatic disease diagnosis from multilingual clinical text, where the input is a free-text symptom description and the output is a disease class. In the reported setting, each example is a short natural-language symptom narrative mapped to one of 24 disease categories. The paper frames the central problem as one of multilingual clinical NLP under low-resource conditions: annotated medical text is limited in Hindi and Bengali, linguistic structure varies across languages, domain-specific medical vocabulary transfers imperfectly, and translation noise can distort symptom descriptions (Sar et al., 24 Sep 2025).

The framework is explicitly organized around zero-shot cross-lingual diagnosis. In this setting, the model is trained primarily on a higher-resource language—stated as mainly English—and then evaluated on Hindi and Bengali without language-specific fine-tuning. This makes SwasthLLM distinct from monolingual diagnostic systems and from approaches that normalize low-resource inputs through translation into English. The paper presents the model as a single multilingual framework rather than a family of language-specific classifiers (Sar et al., 24 Sep 2025).

Related diagnostic systems in health NLP illuminate the problem space that SwasthLLM inhabits. Health-LLM is presented as a personalized retrieval-augmented disease prediction system that converts patient narratives into structured question-based feature scores and feeds them to XGBoost, reaching 0.833 accuracy and 0.762 F1 in its reported setting (Yu et al., 2024). HeLM instead grounds a frozen LLM in individual-specific multimodal data, using modality encoders that map tabular and time-series inputs into the LLM token embedding space; for asthma, it reports AUROC 0.75 with tabular and spirogram data compared with 0.49 when only tabular data are used (Belyaeva et al., 2023). These adjacent systems suggest that SwasthLLM belongs to a broader class of health-LLMs, but its specific contribution is cross-lingual clinical-text diagnosis rather than retrieval-grounded question scoring or multimodal patient-state grounding.

2. Architecture and representation learning

The backbone of SwasthLLM is XLM-RoBERTa-base, selected for cross-lingual representation learning. The encoder output for an input sequence $x^{(l)}$ in language $l \in \{en, hi, bn\}$ is defined as:

$$H^{(l)} = \mathcal{E}_{\theta}(x^{(l)}) = \{h_1, h_2, \dots, h_n\}, \quad h_i \in \mathbb{R}^d$$

with reported encoder settings $d = 768$, max sequence length 128, vocabulary size 250,002, and encoder dropout 0.1 (Sar et al., 24 Sep 2025).

A central component is the language-aware attention mechanism, which assigns higher weights to medically salient tokens despite cross-lingual differences in syntax or token position. The paper gives the attention equations as:

$$\begin{array}{c} e_i = v^{\top} \tan h(W_ah_i + b_a) \ \alpha_i = \frac{\exp (e_i)}{\sum_{j=1}^n \exp (e_j)} \ z = \sum_{i=1}^n \alpha_i h_i \end{array}$$

where $z$ is the attended sentence representation. The implementation table specifies hidden dimension $d' = 256$, Tanh activation, scalar token scores, and attention dropout 0.1 (Sar et al., 24 Sep 2025).

The disease classifier is a linear head over the attended representation:

$\hat{y} = softmax (W_cz + b_c)$

with standard cross-entropy loss:

$$\mathcal{L}_{cls} = - \sum^{|\mathcal{Y}|}_{i=1} y_i \log (\hat{y}_i)$$

for 24 disease classes (Sar et al., 24 Sep 2025).

Cross-lingual alignment is enforced through a Siamese contrastive learning module operating on the encoder’s [CLS] representation. A projection head maps this into a shared latent space:

$$z^{(l)} = MLP_{\phi}(h^{(l)}_{[CLS]}), \quad l \in \{en, hi, bn\}$$

The projection MLP is reported as Linear + ReLU + Linear, with projection dimension 128 (Sar et al., 24 Sep 2025). Contrastive alignment is optimized with an NT-Xent / InfoNCE-style loss:

$l \in \{en, hi, bn\}$0

using cosine similarity and temperature $l \in \{en, hi, bn\}$1 (Sar et al., 24 Sep 2025).

A second alignment term, the translation consistency module, minimizes Euclidean distance between aligned multilingual embeddings:

$l \in \{en, hi, bn\}$2

This directly encourages language-invariant medical representations for translated symptom descriptions (Sar et al., 24 Sep 2025).

3. Training pipeline, objectives, and data regime

SwasthLLM is trained in four phases, with the stated goal of emphasizing representation alignment before task-specific fine-tuning. The first phase is contrastive pretraining, which aligns multilingual symptom descriptions in latent space. The second phase is supervised classification, anchored primarily in English disease labels. The third phase performs joint multi-task training, and the fourth applies MAML to prepare the system for rapid adaptation to unseen languages or tasks with minimal data (Sar et al., 24 Sep 2025).

The joint objective is given as:

$l \in \{en, hi, bn\}$3

with reported weights $l \in \{en, hi, bn\}$4, $l \in \{en, hi, bn\}$5, and $l \in \{en, hi, bn\}$6 (Sar et al., 24 Sep 2025).

The MAML component is formalized through an inner-loop update,

$l \in \{en, hi, bn\}$7

and an outer-loop meta-update,

$l \in \{en, hi, bn\}$8

with inner-loop learning rate $l \in \{en, hi, bn\}$9, meta-learning rate $$H^{(l)} = \mathcal{E}_{\theta}(x^{(l)}) = \{h_1, h_2, \dots, h_n\}, \quad h_i \in \mathbb{R}^d$$0, 5 inner updates, and meta batch size 4 tasks (Sar et al., 24 Sep 2025).

The dataset used in the paper is a parallel multilingual medical text classification dataset of 1200 samples covering 24 disease classes, with approximately 50 samples per class. Each record includes an English symptom description, a direct Hindi translation, a direct Bengali translation, and a disease label. Preprocessing includes punctuation cleanup, whitespace normalization, Unicode normalization for Hindi and Bengali, tokenization with the XLM-RoBERTa tokenizer, and formation of aligned multilingual triplets (Sar et al., 24 Sep 2025).

The train/validation/test regime is described qualitatively rather than numerically: the training set consists primarily of English samples, the validation set is a balanced mix of English, Hindi, and Bengali, and the test set evaluates both supervised English performance and zero-shot Hindi/Bengali transfer. The paper states that stratified sampling preserves class balance across splits (Sar et al., 24 Sep 2025).

The baseline suite is broad. It includes multilingual encoders such as mBERT, IndicBERT, XLM-RoBERTa-base, XLM-RoBERTa-large, mDeBERTa-v3-base, and mT5 variants; medical-domain models such as BioBERT, ClinicalBERT, and MedBERT; and larger instruction-tuned models including MedAlpaca-7B, LLaMA-2-7B, Falcon-7B, and MPT-7B (Sar et al., 24 Sep 2025).

4. Empirical performance and ablation structure

The reported headline results are strong in both supervised and zero-shot settings. In supervised testing, SwasthLLM achieves 97.22% test accuracy and 97.17% F1-score. In zero-shot cross-lingual evaluation, it reports 92.78% accuracy on Hindi and 73.33% accuracy on Bengali. The comparative table further reports Avg F1-score 87.21%, Precision 88.42%, and Recall 86.18%, with SwasthLLM outperforming the listed baselines in the paper’s experimental setup (Sar et al., 24 Sep 2025).

Setting	Metric	Reported result
Supervised	Test accuracy	97.22%
Supervised	F1-score	97.17%
Zero-shot Hindi	Accuracy	92.78%
Zero-shot Bengali	Accuracy	73.33%
Cross-lingual summary	Avg F1-score	87.21%

The paper’s ablation study concentrates on the Bengali zero-shot setting, identified as the hardest condition. A plain Base XLM-R model yields 78.9% accuracy and 76.6% F1. Adding contrastive learning raises this to 83.4% accuracy and 81.2% F1. Adding multi-task learning further raises performance to 85.9% accuracy and 84.0% F1. The full system with contrastive learning + multi-task learning + MAML reaches 87.6% accuracy and 86.2% F1. The paper interprets these increments as evidence that contrastive learning contributes the largest single gain, while multi-task learning and meta-learning add further improvements (Sar et al., 24 Sep 2025).

The reported error analysis states that most Hindi and Bengali misclassifications occur between semantically overlapping diseases, especially tuberculosis vs. pneumonia and dengue vs. typhoid. Relative to baselines, SwasthLLM is reported to have 30% fewer false negatives than mBERT and 22% fewer false positives than mDeBERTa, which the paper presents as clinically relevant because both under-diagnosis and over-prediction matter in diagnostic support (Sar et al., 24 Sep 2025).

The model’s zero-shot asymmetry across languages is also notable. The paper attributes the more difficult Bengali setting to richer morphology, lower resource availability, and weaker alignment for rare idiomatic expressions and culturally specific symptom descriptions. This does not negate the reported performance gain, but it places clear limits on the current degree of language invariance (Sar et al., 24 Sep 2025).

5. Position within Indian healthcare LLM research

Although SwasthLLM itself is a diagnostic classifier rather than a chatbot, related Indian healthcare LLM work shows the broader socio-technical environment into which such a system would be deployed. A reproductive-health chatbot study conducted with underserved women in Mumbai argues that “good medical answers” are insufficient unless the system is also culturally sensitive, linguistically usable, medically grounded, and practically deployable. Its final system uses Hinglish, retrieval-augmented generation (RAG), a localization module, and explicit safety instructions such as avoiding medication prescription and routing to telehealth when appropriate. The same paper proposes a four-layer framework—societal, regional, community, and individual—for culturally sensitive LLM design in healthcare (Deva et al., 21 Feb 2025).

A second line of work, ASHABot, describes a WhatsApp-based, retrieval-grounded, experts-in-the-loop assistant for Indian community health workers. It is explicitly positioned as a supplemental, fallible support tool rather than a replacement for supervisors, and it emphasizes selective escalation, consensus-based human fallback, and local knowledge-base curation (Ramjee et al., 2024). This is highly relevant because it frames the operational role of a health LLM within existing care hierarchies rather than as an autonomous clinician.

Evaluation methodology has likewise moved toward community-grounded assessment. Samiksha proposes a community-driven evaluation pipeline co-created with civil-society organizations and community members in Hindi, Kannada, and Malayalam. Its findings are especially pertinent for SwasthLLM because they show that automated judges align only weakly with human community evaluators in multilingual healthcare-chatbot settings: the paper reports Pearson $$H^{(l)} = \mathcal{E}_{\theta}(x^{(l)}) = \{h_1, h_2, \dots, h_n\}, \quad h_i \in \mathbb{R}^d$$1 between two LLM judges, but only $$H^{(l)} = \mathcal{E}_{\theta}(x^{(l)}) = \{h_1, h_2, \dots, h_n\}, \quad h_i \in \mathbb{R}^d$$2 between LLM judges and humans (Hamna et al., 29 Sep 2025).

At the infrastructure layer, privacy-preserving Indian clinical NLP remains an open prerequisite. Work on Indian clinical discharge summaries shows that de-identification models trained on public non-Indian corpora generalize poorly to Indian notes, while institution-specific synthetic data can raise de-identification performance on a real Indian test set to micro F1 $$H^{(l)} = \mathcal{E}_{\theta}(x^{(l)}) = \{h_1, h_2, \dots, h_n\}, \quad h_i \in \mathbb{R}^d$$3 in the reported setting (Singh et al., 2024). This suggests that any broader SwasthLLM deployment over real clinical data would require institutionally valid privacy pipelines rather than imported de-identification components.

Taken together, these adjacent papers indicate that SwasthLLM solves one important problem—cross-lingual diagnosis from clinical text—but that deployment-grade healthcare LLM systems in India also require cultural grounding, community-centered evaluation, privacy-preserving data curation, and human workflow integration (Deva et al., 21 Feb 2025).

6. Limitations, misconceptions, and prospective extensions

SwasthLLM’s main limitations are explicitly stated in the paper. Bengali remains substantially harder than Hindi; semantically overlapping diseases remain confusable; the language-aware attention mechanism can dilute focus in long-form, multi-comorbidity narratives; the translation module can be weakened by rare idiomatic or culturally specific expressions; and the MAML component increases training time by about 25%. The experimental basis is also narrow: only 1200 samples, 3 languages, and 24 diseases are used, and the paper does not provide a detailed ethics, privacy, or deployment-governance section (Sar et al., 24 Sep 2025).

A common misconception would be to treat SwasthLLM as a general-purpose multilingual clinician. The paper does not support that reading. Its evaluated function is disease classification from symptom narratives, and the broader safety caveat in the technical summary is that it is best understood as a decision-support system, not as a replacement for clinicians (Sar et al., 24 Sep 2025). A plausible implication is that the model’s current strength lies in representation alignment and zero-shot transfer, not in autonomous clinical action.

Future extensions are also visible in adjacent research. Health-LLM for wearable sensor prediction reports that explicit context enhancement can yield up to 23.8% improvement in performance, suggesting that symptom text alone may be only one part of a richer health-state representation (Kim et al., 2024). HeLM shows that multimodal LLM grounding can use clinical tabular data and physiological time series, reporting AUROC 0.75 for asthma with tabular and spirogram inputs compared with 0.49 from tabular-only input (Belyaeva et al., 2023). Guardian Angel, a low-cost wearable paired with a WhatsApp-based LLM agent, reports 100% availability under field noise in its waveform interpretation benchmark and frames the LLM’s role as translating raw physiological data into understandable guidance in constrained settings (Ahsan et al., 9 Feb 2026). These results suggest a plausible future path in which a cross-lingual diagnostic backbone like SwasthLLM is integrated with multimodal patient data, chat-based interfaces, and LMIC-oriented delivery channels.

What remains unresolved is not the existence of a promising architecture, but the gap between a research-grade multilingual classifier and a clinically reliable, socially grounded health system. In that sense, SwasthLLM is best understood as a significant contribution to cross-lingual medical diagnosis using contrastive representations, and as one component of a broader emerging stack for multilingual healthcare AI rather than a complete endpoint in itself (Sar et al., 24 Sep 2025).