Tiantong Lu (Nathan)

Model comparisons

I compare ECG waveform models with logistic-regression baselines trained on clinical features. Feature-ablation experiments test how input choices affect the comparison. Whether the two sources provide complementary information remains an open research question.

Clinical data engineering

The inference count refers to ECG records, not unique patients or the final linked cohort. My linkage and cohort workflows connect model predictions with echocardiography and subsequent hospital observations for retrospective evaluation.

Outcome definitions and phenotyping

The outcome analyses include one-year ventricular tachycardia/fibrillation and newly documented ejection fraction of 35% or less over 1–12 months. Prior ventricular dysfunction, incomplete follow-up and death as a competing risk affect how these endpoints can be interpreted.

I also develop local LLM-assisted phenotyping of discharge summaries to distinguish sustained VT/VF from nonsustained or historical diagnoses.

Current focus

I am testing how clinical-feature choices and phenotype definitions affect cardiac-risk discrimination in retrospective evaluation.

Compositional training data

Labeled tablet images are limited and vary across historical periods and collections. I compose sign crops into line images using corpus text as the target sequence, creating a deduplicated training pool. Evaluation-tablet texts are excluded from synthetic pretraining.

Token-matched controls

Within one CTC recognizer, I designed token-matched single-sign, shuffled-line and attested-line comparisons. Holding the token budget fixed tests training format beyond supervision quantity. Shuffling changes both order and co-occurrence, so these diagnostic comparisons do not isolate a causal contribution from word order alone.

Results and evaluation scope

The reported reduction compares models with and without the compositional synthetic-pretraining stage, followed by adaptation on 100 real tablets. For example, cross-period SVTRv2* sign error fell from 90.8% to 38.4% (three-seed means). Sign error rate counts insertions, deletions and substitutions relative to the reference sign sequence; lower is better.

This measures recognition of reconstructed lines with known sign locations and a fixed vocabulary, not end-to-end tablet transcription. The separate token-matched controls investigate why training formats differ; they do not causally decompose the full pretraining gain.

From recognition to a transcription system

I also develop a modular tablet-transcription pipeline connecting sign detection, visual recognition, language decoding and sequence verification. Separating the stages makes their individual errors and contributions easier to investigate.

For TabletVLM, I contributed synthetic-line supervision and comparative model evaluation to the team’s end-to-end tablet-transcription study.

Trace diagnostics and scoring

The Python tools inspect AB1 signal channels, peaks, basecalls and quality scores. Reference-based scoring measures sequence errors, ambiguous bases and called length. These diagnostics help identify where basecalling and trimming choices affect an output sequence.

Matched-input comparison

The planned evaluation will run TraceTuner and PeakTrace on identical AB1 inputs and compare sequence errors, ambiguity, trimming and reliable read length against a reference. Its error analysis will guide subsequent signal-processing or learning-based development.