arXiv subjects

Raunaq Malhotra

Publications and source records attributed to Raunaq Malhotra.

4 recordsLinked to original sources

arXiv2026

Training a Large Language Model for Medical Coding Using Privacy-Preserving Synthetic Clinical Data

Improving the accuracy and reliability of medical coding reduces clinician burnout and supports revenue cycle processes, freeing providers to focus more on patient care. However, automating the assignment of ICD-10-CM and CPT codes from clinical documentation remains a challenge due to heterogeneous records, nuanced coding guidelines, and long-tail distributions. Large language models have been proposed to help or automate specific medical coding tasks. However, foundation models are not explicitly trained for medical coding and zero-shot coding has yielded poor results. We investigate whether a modern open-weight foundation model can be adapted for an expert-level medical coding task using privacy-preserving synthetic training data derived from electronic health records. We fine-tune Llama 3-70B on pairs of clinical notes and gold codes generated from EHR-grounded templates and coding policies, then evaluate exact-code prediction for ICD-10-CM and CPT. A zero-shot baseline with the unadapted model achieved an F1 score of 0.18 for exact code match. After fine-tuning on the synthetic corpus, exact-match F1 exceeded 0.70, representing a large absolute gain across both code systems. Notably, performance remained high on complex categories that often require multi-step clinical reasoning and code composition, including Advanced Illness and Frailty classes, and the model retained its performance on medical comprehension tasks. These results indicate that synthetic, policy-aware data can efficiently teach a general-purpose large language model to support precise medical coding without exposing protected health information. The approach offers a practical path for training coding agents safely and iteratively on specific tasks that represent real-world populations.

cs.CL↗

arXiv2016

Maximum Likelihood de novo reconstruction of viral populations using paired end sequencing data

We present MLEHaplo, a maximum likelihood de novo assembly algorithm for reconstructing viral haplotypes in a virus population from paired-end next generation sequencing (NGS) data. Using the pairing information of reads in our proposed Viral Path Reconstruction Algorithm (ViPRA), we generate a small subset of paths from a De Bruijn graph of reads that serve as candidate paths for true viral haplotypes. Our proposed method MLEHaplo then generates a maximum likelihood estimate of the viral population using the paths reconstructed by ViPRA. We evaluate and compare MLEHaplo on simulated datasets of 1200 base pairs at different sequence coverage, on HCV strains with sequencing errors, and on a lab mixture of five HIV-1 strains. MLEHaplo reconstructs full length viral haplotypes having a 100% sequence identity to the true viral haplotypes in most of the small genome simulated viral populations at 250x sequencing coverage. While reference based methods either under-estimate or over-estimate the viral haplotypes, MLEHaplo limits the over-estimation to 3 times the size of true viral haplotypes, reconstructs the full phylogeny in the HCV to greater than 99% sequencing identity and captures more sequencing variation for the HIV-1 strains dataset compared to their known consensus sequences.

q-bio.PE↗

arXiv2016

A frame-based representation of genomic sequences for removing errors and rare variant detection in NGS data

We propose a frame-based representation of k-mers for detecting sequencing errors and rare variants in next generation sequencing data obtained from populations of closely related genomes. Frames are sets of non-orthogonal basis functions, traditionally used in signal processing for noise removal. We define a frame for genomes and sequenced reads to consist of discrete spatial signals of every k-mer of a given size. We show that each k-mer in the sequenced data can be projected onto multiple frames and these projections are maximized for spatial signals corresponding to the k-mer's substrings. Our proposed classifier, MultiRes, is trained on the projections of k-mers as features used for marking k-mers as erroneous or true variations in the genome. We evaluate MultiRes on simulated and real viral population datasets and compare it to other error correction methods known in the literature. MultiRes has 4 to 500 times less false positives k-mer predictions compared to other methods, essential for accurate estimation of viral population diversity and their de-novo assembly. It has high recall of the true k-mers, comparable to other error correction methods. MultiRes also has greater than 95% recall for detecting single nucleotide polymorphisms (SNPs), fewer false positive SNPs, while detecting higher number of rare variants compared to other variant calling methods for viral populations. The software is freely available from the GitHub link (https://github.com/raunaq-m/MultiRes).

q-bio.GN↗

arXiv2016

Clustering pipeline for determining consensus sequences in targeted next-generation sequencing

Analyses of targeted genomic sequencing data from next-generation-sequencing (NGS) technologies typically involves mapping reads to a reference sequence or clustering reads. For a number of species a reference genome is not available so the analyses of targeted sequencing data, for example polymorphic structural variation caused by mobile elements is difficult; clustering methods are preferred for such data analysis. Clustering of reads requires a clustering threshold parameter, which is used to compare and group reads. However, determining the optimal clustering threshold for a read dataset is challenging because of different sequence composition, the number of sequences present, and also the amount of sequencing errors in the dataset. High values of the clustering threshold parameter can falsely inflate the number of recovered genomic regions, while low values of clustering threshold can merge reads from distinct regions into a single cluster. Thus, an algorithm that can empirically determine clustering threshold is needed. We propose a pipeline for clustering genomic sequences wherein the clustering threshold is empirically determined from the NGS data. The optimal threshold is decided based on two internal clustering measures which assess clusters for small intra-cluster diameters and large inter-cluster distances. We evaluate the pipeline on two simulated datasets derived from human genome sequence simulating different genomic regions and sequencing depth. The total number of clusters obtained from our pipeline is closer to the actual number of reference sequences when compared to single round of clustering. Also, the number of clusters whose consensus sequence matches a corresponding reference sequence is higher in our pipeline. We observe that the presence of repeat regions affects clustering accuracy.

q-bio.GN↗