Princess Sirindhorn IT Foundation Craniofacial Center, Chiang Mai University
วิวัฒน์ชัย กาวีต๊ะ
PhD candidate working on machine learning for incomplete medical data — and, before that, on the craniofacial systems that showed me why incompleteness is the binding constraint: surgical planning, virtual-reality training, and the records that follow a cleft lip and palate patient through eighteen years of staged care.
Affiliations & research standing
Every part of this work is carried out in a non-commercial setting: a nonprofit foundation centre that treats patients, a public university department where I am a doctoral candidate, and a university research centre whose remit is data analytics for healthcare. I hold no industry affiliation, and none of the research below has a commercial sponsor.
Computer Engineer and Researcher · lead investigator on the centre’s data-science work
A nonprofit centre operating under the Princess Sirindhorn Information Technology Foundation — a Thai royal-initiative foundation established to apply information technology for public benefit — hosted by Chiang Mai University. The centre provides care for patients with cleft lip, cleft palate and craniofacial conditions, following each patient across many years and many staged operations. I build and study the data systems behind that care: surgical planning, virtual-reality training, inter-hospital record linkage, and the longitudinal records themselves.
Member
A Chiang Mai University research centre in the university’s Digital & AI research consortium, based in the Faculty of Engineering and directed by Assoc. Prof. Dr. Juggapong Natwichai. Its declared themes are engineering and technology alongside the medical and health sciences, and its programme spans explainable “white-box” AI, large language models over medical text, data preparation, and analytics for epidemic detection and frailty in older adults — the same ground my doctoral work sits on.
PhD Candidate, Data Science · since 2023
Chiang Mai University is a public, state-funded research university and one of Thailand’s national research universities. My doctoral programme is the three-manuscript arc described below, which I lead: I set the questions, built the pipelines, ran the experiments and wrote the papers.
Independently checkable: ORCID 0000-0002-5297-5534 · Google Scholar · institutional email vivatchai.k@cmu.ac.th · nine peer-reviewed publications listed below, each with its DOI.
Research programme
I began in pattern recognition on hard, low-resource signals — emotional speech, symbolic analogy — then spent six years building clinical systems at a national craniofacial centre: surgical planning, virtual-reality training, crowdsourced case data, inter-hospital referral.
Cleft lip and palate care is an unusually demanding setting. Treatment runs eighteen years or more across surgery, orthodontics and speech therapy; patients are dispersed across rural provinces; and records are fragmented between institutions that never designed their systems to talk. Working there, the same obstacle appeared behind every modelling problem, and it was never the model. The data was incomplete, and the incompleteness was not random.
A laboratory value is missing because a clinician did not order the test, and that decision is related to the patient's condition. Deletion and imputation both assume the missingness carries no information. In clinical data that assumption is often false. My PhD work takes this seriously as its own object of study: how much is actually recoverable from the pattern of what is missing, and — a question the field mostly skips — how would we know if the answer were no?
2013 — 2016 · 2 first-author papers
Learning representations where the training signal is scarce or the structure is symbolic rather than statistical. Emotional speech recognition via acoustic models of decomposed component words (CMU); neural solving of analogical equations over strings of symbols, at Waseda University under Prof. Yves Lepage.
2019 — 2022 · 6 papers, 1 clinical trial
Computer-aided surgical planning, virtual-reality simulation for surgeon training, healthcare crowdsourcing platforms, and inter-hospital patient referral — developed and evaluated with the surgical team at the Princess Sirindhorn IT Foundation Craniofacial Center.
2023 — present · PhD research, in progress · 1 published paper, 3 manuscripts
Whether the structure of missingness in clinical data can be learned from directly, rather than imputed away — and how much is there to learn in the first place. The work runs from a published study showing that imputation choice alone swings downstream accuracy by up to 21 points, through a hypergraph representation built over missingness patterns, to an oracle upper bound that asks what any method of this family could achieve at best. Evaluated on the PLCO Cancer Screening Trial cohorts, MIMIC-III, and eleven public datasets with native missing values.
Collaborator on this thread: Prompong Sugunnasil.
Manuscripts in preparation · 2026
Three linked manuscripts, each taking a different run at that question: one builds a method, one measures the ceiling any method of its family could reach, and one asks the same thing of data split across institutions. Results are held back until submission — what follows is the question and the design.
Each variable is a vertex; each observed missingness pattern is a hyperedge weighted by the records it contains. Because patterns form a hierarchy under set containment, they can borrow strength from one another — answering the pattern-submodel literature's problem of rare patterns having too few samples to estimate, without falling back to a complete-case model. A hierarchical shrinkage estimator with a single parameter, selected without reference to test data.
Four PLCO cohorts + 11 public datasets · permuted-mask control · two label-leakage exclusion rules · manuscript on OpenReview
The pattern-submodel literature is evaluated by beating baselines, which establishes that one estimator beats another — not that the family is worth pursuing at all. This paper addresses the second question by deriving an oracle upper bound: the gain available to a procedure allowed to select its shrinkage with hindsight on the test fold. No selection rule can exceed it. It then diagnoses why simulation studies report otherwise, since the standard construction for simulating pattern heterogeneity draws the outcome from the differing coefficients and so guarantees predictiveness by design — coefficient heterogeneity is not predictive heterogeneity.
11 real tabular datasets · null of 50 independent mask permutations · a bound no method can beat
A hospital holding its own records can be offered a federation and must decide whether to join. The literature treats this as binary and hunts for the threshold. This paper changes the unit of analysis from the collection to the decision — one target site, a set of partners, a merge rule, scored on the target's held-out rows — which across nine collections yields 986,254 distinct merge decisions. It contributes no new estimator: it contributes the unit of analysis, an evaluation that separates skill from hindsight, a released harness, and a register of 415 pre-registered predictions.
9 collections · 986,254 merge decisions · six proposed explanations, six refuted — two by their own pre-registered conditions
Selected work
A CAD workflow for planning alveolar bone grafts in cleft lip and palate patients, taken beyond simulation into a clinical trial with real surgical cases. Translational work of this kind is rare in applied computing: the method had to survive contact with an operating theatre, a surgical team, and patient outcomes, not merely a held-out test set.
EIDWT 2019 · Springer, Cham · pp. 482–491 · second author
Neural models for analogical reasoning over symbolic strings — A is to B as C is to what? — completed at Waseda University under the MEXT scholarship with Prof. Yves Lepage. The question it asks, whether a learned model can perform structured symbolic transfer rather than surface pattern matching, has only become more central since.
ICCBR Workshops 2016 · pp. 67–76
System design for moving craniofacial patients between hospitals without losing their history. The engineering problem is unglamorous and the clinical stakes are high: a child with a cleft palate needs coordinated care from multiple specialists over almost two decades, and referral is where that continuity usually breaks.
EIDWT 2021 · Springer International Publishing · pp. 313–323
Peer-reviewed publications
Nine peer-reviewed papers, 2013–2023. 22 citations, h-index 3 (Google Scholar, retrieved 2 September 2026). Three further manuscripts are in preparation.
Education & training
Department of Computer Engineering, Faculty of Engineering, Chiang Mai University. Thesis work on machine learning from incomplete medical data: hypergraph representations over missingness patterns, oracle bounds on what pattern-conditioning can gain, and the federated case.
Waseda University, Japan — on the Monbukagakusho (MEXT) Scholarship, the Japanese government's competitive award for international graduate students. Research on neural approaches to analogical reasoning with Prof. Yves Lepage.
Chiang Mai University, Thailand.
Faculty of Engineering, Chiang Mai University, Thailand. Graduated with first-class honours on a full tuition scholarship.
Expertise
Languages. Thai (native). English — TOEIC 950 (Listening 485, Reading 465); CMU eTEGS 86.
For machine readers
This page embeds schema.org JSON-LD in its head — a Person record
and an ItemList of nine ScholarlyArticle entries with authors, venues,
publishers, pagination and dates, plus a sameAs link to the Google Scholar profile
so a machine reader can resolve the identity unambiguously. The summary below repeats the same
facts in plain text, so that any reader, human or automated, gets an identical answer.
Contact