QUESTION I EXPLORE
In medicine, a conclusion is only as good as its evidence. Looking back, my research journey has been a search for exactly that — it started with a question about my grandmother's diagnosis, and it led me to a question about AI itself: when a model says it works, how do we know it will work for real patients?
Cleaning the Data AI Learns From
Co-Author (09-12/2025)
To build AI for healthcare, I needed to understand what real medical data looks like. I joined a study on "Building and Standardizing CT Image Datasets for AI-Assisted Lung Nodule Detection" as a co-author, under the guidance of Assoc. Prof. Dr. Nguyễn Hữu Phát (Hanoi University of Science and Technology).
I collected, organized, and reviewed chest CT images, and helped standardize and structure the data so a nodule-detection model could train reliably. It sounds simple, but it took the most time: one batch of images in the wrong format is enough to teach a whole model the wrong thing.
My biggest lesson: most of AI in medicine isn't the model, it's the data. The work was accepted at the 6th International Conference on Intelligent Systems and Networks (ICISN) and the 28th National Conference on Electronics, Communications and Information Technology (REV-ECIT 2025).
Do the Models Tell the Truth?
Author (03/2026 - Present)
Once I had worked with real data, I started to notice that many medical AI papers train on synthetic data. I wanted to know if that's safe.
In my study, "Model Rankings from a Synthetic Multimodal Lung-Cancer Benchmark Do Not Transfer to Real Patients: An Equal-Budget Evaluation with External Validation on TCGA" I compared machine-learning models for diagnosis, treatment recommendation, and survival prediction on a 10,000-patient synthetic benchmark. Every model got the same tuning budget and was tested with calibration analysis, ablation studies, SHAP, bootstrap confidence intervals, and multiple random seeds. Then I externally validated the survival models on 978 real patients from TCGA-LUAD/LUSC.
The answer was no: the model that ranked best on synthetic data was not reliably the best on real patients. It was my most humbling result, and the one that taught me the most about honest science.
THE INTERNSHIP
FPT Smart Cloud
Intern (07-09/2026)
If my Lung Progress Tracker taught me to turn patients’ symptoms into data, FPT Smart Cloud taught me the other direction: turning data back into a voice. In summer 2026, I interned in the Text-to-Speech department, three sessions a week.
I studied how speech synthesis models are trained and contributed to a project building an AI voice agent that answers reception calls for FPT Corporation’s entire office network in place of a human receptionist.
What I realized: moving a model from a notebook into something thousands of people use every day is much harder than training it. And a good AI voice can’t just be correct; it has to make the listener feel respected. I kept thinking of the elderly patients I met at mountain clinics in Hoa Binh and Pu Luong. So after the internship, I added voice chat to my Lung Progress Tracker, letting patients describe their breathing out loud instead of typing it. This time, the data could talk back.
Understanding the Disease
Author (01-10/2026)
My first research paper began with wanting to understand my grandmother's diagnosis.
In my literature review, "Lung Cancer in Vietnam and Worldwide: Epidemiology, Risk Factors, Diagnosis, Conventional and Emerging Treatment Strategies" I studied lung cancer's burden in Vietnam and globally; environmental, occupational, behavioral, and genetic risk factors; screening and diagnosis from low-dose CT to molecular biomarkers; and treatments from chemotherapy to targeted therapy for EGFR, ALK, KRAS, and HER2 mutations.
The chapter I care about most is on Vietnam's gaps: late detection, limited access to molecular testing, and the cost of treatment. Those gaps are exactly the problems my apps try to address.