My paper summaries

Using synthetic data to improve disease classification

Enhancing Disease Classification in Emergency Medical Scenarios Through Synthetic Data: A Comparative Study — Kabir, Nayem & Tomforde.

Does adding computer-generated patient records actually make a diagnosis classifier better — and by how much?

The problem. A model that predicts a patient’s condition learns from labelled examples. In emergency medicine those are hard to come by: datasets are small, cannot be shared for privacy reasons, and are dominated by a few common presentations while the conditions that most need reliable detection are rare. Trained on data like that, a classifier becomes good at the common cases and unreliable exactly where it matters.

What I did. I took a real emergency dataset, generated additional synthetic patient records with several different generative models, and mixed those into the training data in controlled proportions. Then I re-trained the same classifier under each setting and compared the results — paying particular attention to the rare, under-represented conditions rather than just the overall average.

What came out. Synthetic data can give a real improvement, especially for the minority classes, but it is not a free win: the size of the gain depends heavily on which generator is used and how much synthetic data is blended in. Some combinations helped little or slightly hurt. The value is in choosing the method deliberately, not in adding synthetic rows for their own sake.

Why it matters. It is a practical way to strengthen emergency-care models without waiting years for more real cases or exposing patient data — and it shows which generation approach is worth the effort.

In short: synthetic patients help a diagnosis classifier — if you pick the right generator.

Reference: Kabir, M. F., Nayem, M. M. I., & Tomforde, S. (2026). Enhancing disease classification in emergency medical scenarios through synthetic data: a comparative study. In Agents and Artificial Intelligence, Lecture Notes in Computer Science, Springer Nature Switzerland, pp. 286–308. DOI: 10.1007/978-3-032-25035-3_14.

Read the paper →

A diagnosis & treatment assistant for TeleNotArzt care

Diagnosis and Treatment Support as an AI-Driven Assistant System for TeleNotary Emergency Care — Kabir, Piliuk & Tomforde.

The core of my PhD: turning a half-complete, real-time patient record into a ranked list of what is likely wrong and what to do next.

The context. In Germany’s TeleNotArzt system a physician joins an ambulance call remotely and supports the paramedics on scene. They have to judge the situation, decide on treatment, and say which measurements to take — often within a minute or two, from information that is still arriving and never complete.

What I did. I built a deep model that reads three kinds of data at once: time-series vital signs (heart rate, blood pressure over time), tabular clinical fields (age, known conditions, first observations), and free-text notes typed during the call. Each stream is encoded in its own way and then combined, so the model can use whatever is available at that moment. It returns three ranked shortlists — the most probable diagnoses, the treatments worth considering, and the measurements that would most reduce the remaining uncertainty — rather than a single hard answer.

How it is judged. Ranked suggestions need ranked metrics. I used recall@k (is the correct item in the top k suggestions?), macro- and weighted AUC (does it separate classes well, including the rare ones?), and a coverage ratio that checks the recommendations are complete — that nothing important is missing — not merely accurate. For the wider set of ranking metrics this kind of system could use, see the evaluation-techniques survey in Related work.

Why it matters. A shortlist a clinician can scan and accept or reject fits the reality of emergency work better than a black-box verdict, and it degrades gracefully when data is missing. This is the system the other four papers feed into.

In short: one multimodal model, three ranked answers — likely diagnoses, candidate treatments, and what to measure next.

Reference: Kabir, M. F., Piliuk, K., & Tomforde, S. (2026). Diagnosis and treatment support as AI-driven assistant system for TeleNotary emergency care. In Proc. 18th International Conference on Agents and Artificial Intelligence — Vol. 5: ICAART, INSTICC, SciTePress, pp. 4651–4660. DOI: 10.5220/0014481000004052.

Read the paper →

WLSTM-GAN: generating realistic medical tables

Synthetic Data Generation for Emergency Medical Systems: A Systematic Comparison of Tabular GAN Extensions — Kabir, Nayem & Tomforde.

If real patient tables cannot be shared, can we build a stand-in that behaves like the real thing without exposing anyone’s record?

The problem. Every project above needs data, and real emergency tables are both scarce and legally hard to move between institutions. A generative model can produce artificial rows — but medical tables are awkward: they mix continuous values (blood pressure, age) with categorical ones (symptom present or absent, sex, transport type), and many table generators handle that mixture badly, blurring categories or dropping rare values.

What I did. I designed WLSTM-GAN, a GAN with a split architecture: one generator containing two specialised LSTM sub-networks, paired with two separate discriminators — one that judges the continuous columns and one that judges the categorical columns. Each data type is therefore learned and critiqued on its own terms. I then ran a systematic comparison against other tabular-GAN variants on three medical datasets, scoring each on fidelity (does the synthetic data reproduce the real statistics and relationships?) and on privacy (can a real record be recovered from the synthetic set?), both with and without an explicit privacy constraint.

What came out. WLSTM-GAN produced data that was closer to the real distribution and harder to trace back to individuals than the baseline generators — the split design pays off precisely because it stops treating a medical table as one uniform block of numbers.

Why it matters. It is the synthetic-data engine the classification and assistant-system work rely on, and a route to sharing emergency-medicine data for research without sharing patients.

In short: give the continuous and categorical columns their own LSTM generator and discriminator, and the synthetic tables come out both more realistic and more private.

Reference: Kabir, M. F., Nayem, M. M. I., & Tomforde, S. (2025). Synthetic data generation for emergency medical systems: a systematic comparison of tabular GAN extensions. In Proc. 17th International Conference on Agents and Artificial Intelligence — Vol. 3: ICAART, INSTICC, SciTePress, pp. 1199–1206. DOI: 10.5220/0013307200003890.

Read the paper →

Filling in missing values in emergency data

A Deep Analysis for Medical Emergency Missing Value Imputation — Kabir & Tomforde.

Before any model can learn from emergency data, someone has to decide what goes in the empty cells — and that choice changes the results.

The problem. The TeleNotArzt dataset is missing far more values than a typical medical dataset, and the gaps are scattered essentially at random (a measurement simply was not taken in time) rather than following a pattern a model could exploit. On top of that the data is strongly imbalanced and full of outliers from genuine rare emergencies. A careless fill-in — replacing every gap with a column average, say — erases exactly the unusual signals that matter most.

What I did. I compared a range of options for reconstructing the missing numeric values: standard statistical and machine-learning imputers, a denoising autoencoder, and a denoising LSTM autoencoder (which can use the order and timing of measurements, not just their values). All were tested on the real, non-public emergency dataset and judged on how closely the reconstructed values matched held-out ground truth.

What came out. The denoising LSTM autoencoder was consistently the most accurate, because it learns the temporal shape of a patient’s measurements rather than treating each value in isolation — and it held up better on the outlier cases.

Why it matters. Imputation is the first step in the pipeline; every later model inherits its errors. Getting it right, especially for rare cases, is a prerequisite for a trustworthy assistant.

In short: for time-stamped emergency measurements, a denoising LSTM autoencoder fills the gaps better than conventional imputation.

Reference: Kabir, M. F., & Tomforde, S. (2024). A deep analysis for medical emergency missing value imputation. In Proc. 16th International Conference on Agents and Artificial Intelligence — Vol. 3: ICAART, INSTICC, SciTePress, pp. 1229–1236. DOI: 10.5220/0012457300003636.

Read the paper →

Deep learning for highly heterogeneous emergency data

A Deep Learning Approach for the Classification of Highly Heterogeneous Data in the Context of Treatment of Medical Emergencies — Aust, …, Kabir et al. (Organic Computing, ch. 10).

The starting question: can one deep-learning pipeline make sense of emergency records where every case looks structurally different?

The problem. A single emergency record can contain numbers, categories, timestamps, and text, recorded by different people to different standards, with large parts absent. Classical models expect tidy, uniform inputs and struggle with this kind of variety.

What I did. As part of a multi-author chapter from the Organic Computing programme, I contributed to a deep-learning pipeline that ingests this heterogeneous data and classifies several targets at once — diagnostic categories, medications, and physiological measurements — so that a TeleNotArzt receives structured suggestions faster.

Why it matters. This was the proof of concept: it showed the heterogeneity of emergency data is workable, and it set up the more focused papers on data quality, synthetic data, and the full assistant system.

In short: the first step — one pipeline that classifies diagnoses, medications, and measurements straight from messy emergency records.

Reference: Aust, T., Cui, H., Botache, D., Decke, J., Frost, E., Gerling, V., Heider, M., Henrichs, E., Hubert, A., Kabir, M. F., Kisselbach, T., Meitz, L., Pilar von Pilchau, W., Oguz Saltık, A., Schönberner, C., Spils, M., Steinberger, M., Kaul, N.-H., & Reichert, H. (2024). A deep learning approach for the classification of highly heterogeneous data in the context of treatment of medical emergencies. In S. Tomforde & C. Krupitzer (Eds.), Organic Computing, ch. 10, kassel university press, pp. 121–135. DOI: 10.17170/kobra-202402269661.

Read the paper →