AI for autism intervention: Developing an integrated AI system for childhood autism that generate evaluation reports, personalized intervention strategies and post-intervention plans based on patient data analysis.

Project 1：LLM-based Clinical Decision Support System (LLM-CDSS) for Autism Spectrum Disorder

• Designed and implemented an LLM-based multi-agent system where specialized agents simulate key clinical roles in ASD intervention through two main functions: conducting assessments and developing intervention plans.
• Designed and conducted a large-scale clinical validation study (n=70) to evaluate the impact of LLM-CDSS on standardizing autism intervention decisions and reducing clinician-dependent variations in treatment planning.
• Implemented a controlled pre-post study with 4,200 randomized cases and expert-validated gold standards (n=5 senior clinicians), demonstrating significant improvements in clinical decision accuracy (17-19%), consistency (25%), and efficiency (54% time reduction, p<0.001).

Project 2：AI-Powered Interactive Story Generation Platform for Children with Autism

• Developed a Stable Diffusion-based model fine-tuned with LoRA, ensuring consistent style and therapeutic value in narratives.
• Implemented a dynamic prompt engineering system with context-aware templates and few-shot examples, integrating Chain-of-Thought reasoning to enhance coherence and educational impact.
• Built a full-stack platform for real-time story generation and feedback, now used exclusively by therapists with over 1,000 interactions, providing engaging, adaptive stories for autistic children.

AI for hematology analyzer: Leveraging machine learning to enhance hematology analyzers, enabling cost-effective and rapid early diagnosis of sepsis through advanced blood cell counting and classification techniques.

Project 1: Advancements in AI-Based Hematological Analysis

• Developed a Class Activation Mapping (CAM)-based feature selection method for blood biomarkers, which streamlined algorithm design and optimization by reducing feature space 80%.

Project 2: Innovative Sepsis Diagnosis Research

• Led research project on novel hematological parameters for early sepsis diagnosis.
• Implemented a targeted genetic algorithm, enabling precise identification of top N blood cell biomarkers.
• Developed a rule-based algorithm for optimal usage of biomarker panels in varied clinical scenarios, surpassing traditional single biomarker limitations.
• Outperformed current diagnostic methods (CRP\PCT) with superior accuracy (Acc 21%↑& AUC 14%↑), faster results, and cost-efficiency.

AI in Surgical Enhancement: Enhancing laparoscopic surgery by mitigating visibility issues due to smoke, and enabling precise tissue targeting in dynamic environments for ultrasound-guided HIFU therapy through advanced predictive modeling.

Project 1: Efficient Model for Medical Image Enhancement

• Led the development of the PFAN model, a lightweight GAN framework integrating CNN and Transformer based on frequency domain for desmoking laparoscopic images.
• Outperformed existing benchmarks in PSNR, SSIM, CIEDE2000, and visual effects.
• Research accepted by PRCV 2023.

Project 2: Spatiotemporal Model for Soft Tissue Motion Prediction

• Developed STU-Net, an Encoder-Decoder framework utilizing spatial and temporal information in HIFU image sequences for improved motion prediction and segmentation.
• Achieved high performance metrics (98.65% BA, 94.32% Dice, 89.26% mIoU) with STU-Net and a prediction speed of 0.23 seconds, enabling real-time treatment adjustments.
• Culminated in a clinically applicable method, recognized with CN Patent ZL202110345245.8.

Data Mining in Social Media: Utilizing machine learning and feature engineering to extract valuable information from social media data, enabling more effective ad targeting by identifying user preferences and trending topics.

Project: Innovative Advertisement Video Recognition Method

• Extracted audiovisual features like spectogram and duration using FFmpeg instead of visual features for advertisement video identification, achieving 92% accuracy and 98% AUC through LightGBM model.
• Scripts still run in production today.

• Built and managed social media platforms catering specifically to college students, amassing over 60,000 active users.
• Built a toolbox that included useful tools like credit calculators and course reminders, helping students with their studies.

Machine Translation: Addressing the complexity of translating concise and allusion-rich ancient Chinese into understandable modern language through advanced neural network technology.

Project: Neural Network for Machine Translation

• Proposed a convolution-based fixed-pattern noise removal method to optimize text recognition of OCR. Method accepted by ICAIBD 2018.
• Build the largest parallel corpus of ancient Chinese and modern Chinese (78,000+ pairs) at that time.
• Proposed a Character-based DBRNN Model for Ancient-Modern Chinese Neural Machine Translation which outperforms state-of-the-art models in terms of BLEU.
• Rated as National Excellent Innovation and Entrepreneurship Project.