Pith. sign in

REVIEW 4 major objections 4 minor 30 references

Developing hybrid mechanistic and data-driven personalized prediction models for platelet dynamics

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a purely data-driven recurrent network, pre-trained on synthetic mechanistic trajectories and fine-tuned per patient, predicts individual platelet counts during chemotherapy better than mechanistic or hybrid models…

desk verdict Solid dense-data result but the sparse-data advantage of mechanistic models is likely a metric artifact; needs revision before the two-part conclusion can be taken at face value. read the letter →

arxiv 2505.21204 v1 pith:RWXJ5QTH submitted 2025-05-27 cs.LG q-bio.QM

Marie Steinacker , Yuri Kheifetz , Markus Scholz This is my paper
classification cs.LGq-bio.QM
keywords plateletdynamicschemotherapy-inducedthrombocytopeniauniversaldifferentialequationsmechanisticmodelsNARXnetworkstransferlearningtimeseriesforecastingpersonalizedmedicine
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks which modeling strategy best predicts an individual patient's platelet counts during multi-cycle chemotherapy: a mechanistic differential-equation model of blood-cell production, a hybrid that augments the mechanistic model with neural-network terms, or a purely data-driven recurrent network. It claims that when a patient has enough dense data (at least three treatment cycles with multiple measurements per cycle), the fully data-driven ARX-GRU model — pre-trained on synthetic trajectories generated by the mechanistic model and fine-tuned on the patient's own data — forecasts subsequent cycles more accurately than any mechanistic or hybrid alternative, with the advantage concentrated in high-risk patients with deep, irregular nadirs. For sparse data, the paper claims mechanistic and hybrid models remain superior or comparable. This matters because chemotherapy-induced thrombocytopenia is highly variable between patients, and choosing the right model per patient could support adaptive, personalized treatment planning.

What carries the argument

The central object is the ARX-GRU model: a nonlinear autoregressive recurrent neural network with exogenous inputs (chemotherapy administration) using gated recurrent units, trained per patient via transfer learning — first pre-trained on synthetic platelet trajectories simulated from the Friberg mechanistic model across virtual therapy scenarios, then fine-tuned on the patient's real measurements. This two-stage procedure lets a purely data-driven model inherit physiological structure from the mechanistic model while retaining flexibility to capture irregular individual dynamics. The hybrid UDE models (UDE-add and UDE-rep) instead embed a small neural network into the feedback term of the Friberg model's stem-cell equation, with non-negativity enforced by a tanh factor, so they are mechanistic models with learnable corrections. The comparison metric is a weighted mean-squared error (SMSE) that tolerates imprecise measurement times by comparing predictions at t−1, t, and t+1.

What would settle it

Re-score the sparse-data scenarios with a nadir-focused metric that only rewards predictions when the platelet count actually crosses a clinical threshold (e.g., below 50×$10^{9}$/L) at the observed nadir time; if the Henrich model's SMSE advantage vanishes or reverses under that metric, the paper's claim that mechanistic models are superior for sparse data is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the best model choice depends on data availability, and in the data-rich regime a purely data-driven recurrent network outperforms mechanistic and hybrid models. Concretely, in the dense 14-day-cycle group (De14) ARX-GRU outperforms all other approaches, with the improvement becoming significant after three training cycles; in the dense 21-day group (De21) it outperforms all others with three or more cycles. In contrast, for sparse data the Henrich mechanistic model is often best, and the UDE hybrid models are comparable to the refined mechanistic models while beating the simpler Friberg model. The paper also claims that this framework generalizes to other hematologic lineages and treatment-related toxicities.

Load-bearing premise

The conclusion that mechanistic and hybrid models are superior in sparse-data scenarios assumes that the sparse measurements capture the clinically relevant nadir phase; the paper itself acknowledges that with sparser data the important nadir phase may not be captured.

Editorial extensions

If this is right

  • For dense data (at least three cycles with multiple measurements per cycle), the ARX-GRU model provides the best held-out-cycle forecasts of platelet nadirs, which is the clinically relevant event.
  • Hybrid UDE models beat the simple Friberg model and match refined mechanistic models without additional biological assumptions, so mechanistic baselines can be improved by learnable corrections.
  • In sparse-data scenarios, mechanistic (especially Henrich) models remain preferable; data-driven models degrade toward average trajectories.
  • The modeling and comparison framework is generalizable to other blood lineages and chemotherapy toxicities, since it only requires repeated measurements of a count plus treatment timing.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The crossover point between data-driven and mechanistic regimes likely depends not only on cycle count but on nadir coverage; clinics that routinely sample during the expected nadir window may see ARX-GRU win with fewer cycles.
  • Because the ARX-GRU model is pre-trained on synthetic Friberg trajectories, the 'purely data-driven' model is not fully agnostic — testing it without mechanistic pretraining would separate the value of the architecture from the value of the mechanistic prior.
  • The SMSE metric rewards point-wise proximity with time tolerance; a clinically oriented extension would evaluate whether predicted nadir depth and timing fall within a safety-relevant threshold, which could change the ranking of models that miss the nadir entirely.
  • The findings suggest an adaptive model-selection rule: use mechanistic or hybrid models for the first two cycles, then switch to ARX-GRU once dense multi-cycle data accumulate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper develops and compares mechanistic, hybrid (UDE), and data-driven (ARX-GRU) models for individualized forecasting of platelet counts during multi-cycle chemotherapy. Using data from 360 NHL-B trial patients split into dense/sparse and 14/21-day cycle groups, the authors train on up to five earlier cycles and evaluate on held-out later cycles under a modified SMSE objective. The main reported findings are that ARX-GRU outperforms all other models in the dense groups after three or more training cycles, while mechanistic or hybrid models are superior in sparse-data scenarios. The authors also provide individual case studies, code, and detailed computational cost reporting.

Significance. The dense-data result is an empirically interesting and potentially useful contribution: a purely data-driven recurrent model with mechanistic pre-training can beat established mechanistic models for held-out-cycle platelet prediction in patients with frequent measurements, especially for irregular, high-toxicity trajectories. The evaluation is genuinely out-of-sample, the cohort is real, and the code is publicly available, which strengthens reproducibility. However, the sparse-data half of the central claim is considerably weaker because the evaluation metric may not measure the clinically relevant nadir phase when only two measurements per cycle are available. If the sparse-data conclusion is reframed or restricted to cycles with observed nadirs, the paper's main contribution would rest on the dense-data comparison, which is better supported.

major comments (4)
  1. [Section 3.4, Conclusion, Eq. (8)] The sparse-data superiority claim for mechanistic/hybrid models is not adequately supported because the SMSE metric (Eq. 8) scores predictions only at observed time points and their immediate neighbors. In the Sp groups, with at most two measurements per cycle, the nadir is usually unobserved, as the paper itself acknowledges in the Conclusion: 'with these sparser data, the important nadir phase of cell counts might not be captured by the data.' A model that produces a smooth population-average trajectory can match the observed early-decline and late-recovery points and receive a good SMSE without ever predicting the true minimum. This directly affects the abstract's claim that 'hybrid and mechanistic models are superior in scenarios with limited or sparse data.' I recommend restricting the sparse-data evaluation to cycles where a nadir is actually observed, or adding a nadir-focused metric (e.g., predicted vs. observed minimum and its timing), and comparing against a trivial baseline that predicts the group average trajectory to show added clinical skill.
  2. [Section 3.1] The patient selection criterion of at least four completed therapy cycles and more than one blood count per cycle introduces a survivorship bias that is not discussed. Patients with early treatment discontinuation due to severe toxicity or poor adherence are excluded, which may underrepresent the very high-risk, irregular trajectories that the paper emphasizes in its abstract. The authors should report how many patients were excluded by each criterion and provide a sensitivity analysis or explicitly limit the generalization claim to patients who can complete at least four cycles.
  3. [Section 3.4, Figure 2] The statistical reporting is incomplete for the number of comparisons made. The one-sided Wilcoxon tests are run at p=0.05 across seven models and five training-cycle settings per panel, yielding 35 comparisons per group with no correction for multiplicity. This makes the 'significantly' statements in the text, such as 'ARX-GRU outperforms all other modeling approaches, which becomes significant after three training cycles,' vulnerable to false positives. Please provide corrected p-values, a pre-specified primary comparison, or a clear justification for why correction is unnecessary.
  4. [Section 3.4] The sparse-data conclusion is also internally mixed: in Sp14 the Henrich model is reported as superior, but in Sp21 the ARX-GRU is reported as superior after three training cycles with the difference to Henrich not significant. Thus the claim that mechanistic and hybrid models are generally superior under sparsity is not uniformly supported by the results as presented. The text should either separate the claims by group and training-cycle count or hedge the conclusion accordingly.
minor comments (4)
  1. [Section 3.5] For patient #4, the text states that the mechanistic and UDE-add models predicted the treatment-delay spike, while ARX-GRU could not, yet then says ARX-GRU predictions of the nadir phases match observed nadirs best. Please clarify which quantity is being prioritized, because these two statements appear to reward different aspects of the trajectories.
  2. [Section 3.3] The SMSE equation uses time indices t-1 and t+1, but the manuscript does not specify the time resolution of the model grid. Please state whether these are calendar days, model time steps, or something else, as this affects how the 0.3 weights should be interpreted.
  3. [Section 3.4] The sentence 'The bad performance of the MS-rev was due to one outlier with particularly bad fitting behavior' is too vague. If one outlier drives the aggregate result, please report the effect of removing that patient or provide the individual SMSE distribution.
  4. [Figure 2 caption] The heatmap color scale makes it difficult to assess the magnitude of SMSE differences from the text alone. Please include a supplementary table with the exact SMSE values and standard errors for each model and training-cycle setting.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central comparisons are out-of-sample predictions evaluated on held-out therapy cycles.

full rationale

The paper's main claims rest on a genuine out-of-sample evaluation. For each patient, mechanistic, hybrid UDE, and ARX-GRU models are fitted using one to five therapy cycles and then scored with the SMSE objective (Eq. 8) on subsequent cycles not used for training. The ARX-GRU pre-training uses synthetic data generated from the Friberg model, and the UDE variants embed Friberg structure, but neither construction makes the held-out prediction equal to a fitted input; the Friberg parameters used for initialization do not determine the test-cycle SMSE values. The self-citations (refs. 18 and 22) motivate the ARX-GRU architecture and transfer-learning scheme, but the reported advantages—ARX-GRU for De14/De21, mechanistic/hybrid for sparse groups—are established by the computed comparisons, not by the citations themselves. The paper's own conclusion that sparse data may not capture nadir phases is a limitation on clinical interpretability, not a circular dependency. No equation reduces a claimed prediction to its own inputs. The evaluation is self-contained against real clinical data, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The models depend on several fitted parameters, including per-patient mechanistic coefficients, neural network weights, the UDE scaling factor, and the SMSE weighting. Domain assumptions are inherited from the Friberg model, the step-function drug effect, log-transformation, the SMSE objective, and the patient selection criteria. No new biological entities are postulated.

free parameters (5)
  • per-patient mechanistic parameters = not reported (estimated per patient)
    kp=ktr=kc, C0, gamma, and the Edrug proportionality constant are fit to each patient's log-platelet data with a penalty toward population means, as described in Section 3.2.
  • UDE scaling factor a = 0.005
    Chosen by hyperparameter optimization to enforce non-negativity of the neural augmentation in Equations (6) and (7).
  • SMSE time-neighborhood weight = 0.3
    Weight assigned to t-1 and t+1 prediction errors in the objective function (Equation 8), chosen based on prior work by the same group.
  • neural network weights and architecture = trained
    Weights of the UDE and ARX-GRU networks are trained on individual patient data; architecture hyperparameters such as layer sizes and learning rates are selected via search ranges given in the supplement.
  • penalty strength for individual parameter regularization = not stated in main text
    Balances individual model fidelity against population consistency in mechanistic model estimation, mentioned in Section 3.2.
assumptions (5)
  • domain assumption The Friberg model equations (1)-(5) adequately describe core platelet dynamics under cytotoxic chemotherapy.
    Used as the mechanistic base for the hybrid models and as a baseline; introduced in Section 2.1.
  • ad hoc to paper Chemotherapy effect can be represented as a step function proportional to drug dosage, nonzero only on treatment days.
    Simplification adopted in Section 2.1 to model Edrug, without explicit validation against detailed pharmacokinetic data.
  • domain assumption Log-transformed platelet counts are the appropriate scale for modeling and error evaluation.
    Stated in Section 3.2 as the basis for learning.
  • ad hoc to paper SMSE (Equation 8) is a valid measure of predictive accuracy for clinical forecasting.
    The authors justify it by uncertain measurement timing, but it alters what 'good prediction' means compared to standard point-wise error.
  • domain assumption Patients with at least four cycles and more than one measurement per cycle are representative of the target clinical population.
    Selection criteria in Section 3.1 exclude patients who discontinued early, likely biasing toward less severe toxicity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Developing hybrid mechanistic and data-driven personalized prediction models for platelet dynamics." pith.science (2026). https://pith.science/paper/RWXJ5QTH

@misc{pith2026250521204,
  author       = {Pith},
  title        = {Pith review of: Developing hybrid mechanistic and data-driven personalized prediction models for platelet dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RWXJ5QTH}},
  note         = {Machine review of arXiv:2505.21204}
}
read the original abstract

Hematotoxicity, drug-induced damage to the blood-forming system, is a frequent side effect of cytotoxic chemotherapy and poses a significant challenge in clinical practice due to its high inter-patient variability and limited predictability. Current mechanistic models often struggle to accurately forecast outcomes for patients with irregular or atypical trajectories. In this study, we develop and compare hybrid mechanistic and data-driven approaches for individualized time series modeling of platelet counts during chemotherapy. We consider hybrid models that combine mechanistic models with neural networks, known as universal differential equations. As a purely data-driven alternative, we utilize a nonlinear autoregressive exogenous model using gated recurrent units as the underlying architecture. These models are evaluated across a range of real patient scenarios, varying in data availability and sparsity, to assess predictive performance. Our findings demonstrate that data-driven methods, when provided with sufficient data, significantly improve prediction accuracy, particularly for high-risk patients with irregular platelet dynamics. This highlights the potential of data-driven approaches in enhancing clinical decision-making. In contrast, hybrid and mechanistic models are superior in scenarios with limited or sparse data. The proposed modeling and comparison framework is generalizable and could be extended to predict other treatment-related toxicities, offering broad applicability in personalized medicine.

Figures

Figures reproduced from arXiv: 2505.21204 by the authors.

Figure 1
Figure 1. Combining mechanistic and data-driven models. Mechanistic models and data-driven [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Average prediction performance per patient group and data scenario. We compare per [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Individual predictions for four selected patients. We present predictions of the Friberg [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 24 canonical work pages

  1. [1]

    Pfreundschuh

    M. Pfreundschuh. Two-weekly or 3-weekly CHOP chemotherapy with or without etoposide for the treatment of young patients with good-prognosis (normal LDH) aggressive lymphomas: results of the NHL-B1 trial of the DSHNHL. Blood, 104(3):626–633, April 2004. ISSN 0006- 4971, 1528-0020. doi: 10.1182/blood-2003-06-2094. URL http://www.bloodjournal. org/cgi/doi/10...

    work page doi:10.1182/blood-2003-06-2094 2004

  2. [2]

    Pfreundschuh

    M. Pfreundschuh. Two-weekly or 3-weekly CHOP chemotherapy with or without etoposide for the treatment of elderly patients with aggressive lymphomas: results of the NHL-B2 trial of the DSHNHL. Blood, 104(3):634–641, April 2004. ISSN 0006-4971, 1528-0020. doi: 10.1182/blood-2003-06-2095. URL http://www.bloodjournal.org/cgi/doi/10.1182/ blood-2003-06-2095 . 9

    work page doi:10.1182/blood-2003-06-2095 2004

  3. [3]

    Dale, and Gary H

    Jeffrey Crawford, David C. Dale, and Gary H. Lyman. Chemotherapy-induced neutropenia: Risks, consequences, and new directions for its management. Cancer, 100(2):228–237, January

    work page

  4. [4]

    Wunderlich, M

    A. Wunderlich, M. Kloess, M. Reiser, C. Rudolph, L. Truemper, S. Bittner, H. Schmalen- berg, R. Schmits, M. Pfreundschuh, and M. Loeffler. Practicability and acute haematological toxicity of 2- and 3-weekly CHOP and CHOEP chemotherapy for aggressive non-Hodgkin’s lymphoma: results from the NHL-B trial of the GermanHigh-Grade Non-Hodgkin’s Lym- phoma Study...

    work page doi:10.1093/annonc/mdg249 2003

  5. [5]

    Friberg, Anja Henningsson, Hugo Maas, Laurent Nguyen, and Mats O

    Lena E. Friberg, Anja Henningsson, Hugo Maas, Laurent Nguyen, and Mats O. Karlsson. Model of Chemotherapy-Induced Myelosuppression With Parameter Consistency Across Drugs. Journal of Clinical Oncology , 20(24):4713–4721, December 2002. ISSN 0732-183X, 1527-

    work page 2002

  6. [6]

    Friberg and Mats O

    Lena E. Friberg and Mats O. Karlsson. Mechanistic models for myelosuppression. Investiga- tional New Drugs, 21(2):183–194, 2003. ISSN 01676997. doi: 10.1023/A:1023573429626. URL http://link.springer.com/10.1023/A:1023573429626

    work page doi:10.1023/a:1023573429626 2003

  7. [7]

    Garrido, Elena Soto, and Iñaki F

    Víctor Mangas-Sanjuan, Núria Buil-Bruna, María J. Garrido, Elena Soto, and Iñaki F. Trocóniz. Semimechanistic Cell-Cycle Type–Based Pharmacokinetic/Pharmacodynamic Model of Chemotherapy-Induced Neutropenic Effects of Diflomotecan under Different Dosing Sched- ules. The Journal of Pharmacology and Experimental Therapeutics , 354(1):55–64, July 2015. ISSN 0...

    work page doi:10.1124/jpet.115.223776 2015

  8. [8]

    Semimechanistic Bone Marrow Exhaus- tion Pharmacokinetic/Pharmacodynamic Model for Chemotherapy-Induced Cumulative Neu- tropenia

    Andrea Henrich, Markus Joerger, Stefanie Kraff, Ulrich Jaehde, Wilhelm Huisinga, Char- lotte Kloft, and Zinnia Patricia Parra-Guillen. Semimechanistic Bone Marrow Exhaus- tion Pharmacokinetic/Pharmacodynamic Model for Chemotherapy-Induced Cumulative Neu- tropenia. Journal of Pharmacology and Experimental Therapeutics , 362(2):347–358, Au- gust 2017. ISSN ...

    work page doi:10.1124/jpet.117.240309 2017

Show all 30 references
  1. [9]

    A biomathematical model of human throm- bopoiesis under chemotherapy

    Markus Scholz, Arnd Gross, and Markus Loeffler. A biomathematical model of human throm- bopoiesis under chemotherapy. Journal of Theoretical Biology , 264(2):287–300, May 2010. ISSN 00225193. doi: 10.1016/j.jtbi.2009.12.032. URL https://linkinghub.elsevier. com/retrieve/pii/S0...

    2010 doi

  2. [10]

    Modeling individual time courses of thrombopoiesis during multi-cyclic chemotherapy

    Yuri Kheifetz and Markus Scholz. Modeling individual time courses of thrombopoiesis during multi-cyclic chemotherapy. PLOS Computational Biology, 15(3):e1006775, March 2019. ISSN 1553-7358. doi: 10.1371/journal.pcbi.1006775. URL https://dx.plos.org/10.1371/ journal.pcbi.1006775

    2019 doi

  3. [11]

    Individual prediction of thrombocytopenia at next chemother- apy cycle: Evaluation of dynamic model performances

    Yuri Kheifetz and Markus Scholz. Individual prediction of thrombocytopenia at next chemother- apy cycle: Evaluation of dynamic model performances. British Journal of Clinical Pharmacol- ogy, page bcp.14722, January 2021. ISSN 0306-5251, 1365-2125. doi: 10.1111/bcp.14722. URL h...

    2021 doi

  4. [12]

    Stewart, and Jimeng Sun

    Edward Choi, Mohammad Taha Bahadori, Andy Schuetz, Walter F. Stewart, and Jimeng Sun. Doctor AI: Predicting Clinical Events via Recurrent Neural Networks. JMLR workshop and conference proceedings, 56:301–318, August 2016. ISSN 1938-7288

    2016

  5. [13]

    Predict- ing Clinical Events by Combining Static and Dynamic Information Using Recurrent Neural Networks

    Cristobal Esteban, Oliver Staeck, Stephan Baier, Yinchong Yang, and V olker Tresp. Predict- ing Clinical Events by Combining Static and Dynamic Information Using Recurrent Neural Networks. pages 93–101, October 2016. doi: 10.1109/ICHI.2016.16

    2016 doi

  6. [14]

    G. Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of Control, Signals and Systems, 2(4):303–314, December 1989. ISSN 1435-568X. doi: 10.1007/ BF02551274. URL https://doi.org/10.1007/BF02551274. 10

    1989 doi

  7. [15]

    Universal Differential Equations for Scientific Machine Learning, November 2021

    Christopher Rackauckas, Yingbo Ma, Julius Martensen, Collin Warner, Kirill Zubov, Rohit Supekar, Dominic Skinner, Ali Ramadhan, and Alan Edelman. Universal Differential Equations for Scientific Machine Learning, November 2021. URL http://arxiv.org/abs/2001. 04385. arXiv:2001.0...

    2021 arXiv

  8. [16]

    Raissi, P

    M. Raissi, P. Perdikaris, and G.E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics , 378:686–707, February 2019. ISSN 00219991...

    2019 doi

  9. [17]

    Heyder, Julian Zabbarov, Pascal Iversen, Simon Witzke, Bernhard Y

    Maximilian Kleissl, Lukas Drews, Benedict B. Heyder, Julian Zabbarov, Pascal Iversen, Simon Witzke, Bernhard Y . Renard, and Katharina Baum. SimbaML: Connecting Mechanistic Models and Machine Learning with Augmented Data. In Krystal Maughan, Rosanne Liu, and Thomas F. Burns, e...

    2023

  10. [18]

    Individual modelling of haematotoxicity with NARX neural networks: A knowledge transfer approach

    Marie Steinacker, Yuri Kheifetz, and Markus Scholz. Individual modelling of haematotoxicity with NARX neural networks: A knowledge transfer approach. Heliyon, 9(7):e17890, July

  11. [19]

    Renard, and Katharina Baum

    Julian Zabbarov, Simon Witzke, Maximilian Kleissl, Pascal Iversen, Bernhard Y . Renard, and Katharina Baum. Optimizing ODE-derived Synthetic Data for Transfer Learning in Dynamical Biological Systems, March 2024. URLhttp://biorxiv.org/lookup/doi/10.1101/2024. 03.25.586390

    2024 doi

  12. [20]

    Data-Driven Dis- covery of Feedback Mechanisms in Acute Myeloid Leukaemia: Alternatives to classical models using Deep Nonlinear Mixed Effect modeling and Symbolic Regression

    Carl Julius Martensen, Niklas Korsbo, Vijay Ivaturi, and Sebastian Sager. Data-Driven Dis- covery of Feedback Mechanisms in Acute Myeloid Leukaemia: Alternatives to classical models using Deep Nonlinear Mixed Effect modeling and Symbolic Regression. bioRxiv, page 2024.06.17.59...

    2024 doi

  13. [21]

    Siegelmann, B.G

    H.T. Siegelmann, B.G. Horne, and C.L. Giles. Computational capabilities of recurrent NARX neural networks. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) , 27(2):208–215, April 1997. ISSN 1083-4419, 1941-0492. doi: 10.1109/3477.558801. URL https://iee...

    1997

  14. [22]

    Predicting chemotherapy-induced thrombotoxicity by NARX neural networks and transfer learning

    Marie Steinacker, Yuri Kheifetz, and Markus Scholz. Predicting chemotherapy-induced thrombotoxicity by NARX neural networks and transfer learning. Journal of Can- cer Research and Clinical Oncology , 150(10):457, October 2024. ISSN 1432-1335. doi: 10.1007/s00432-024-05985-y. U...

    2024 doi

  15. [23]

    Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural Ordinary Differential Equations. In Advances in Neural Information Processing Systems , volume 31. Cur- ran Associates, Inc., 2018. URL https://papers.nips.cc/paper_files/paper/2018/ hash/69386f6...

    2018

  16. [24]

    Non- Negative Universal Differential Equations With Applications in Systems Biology

    Maren Philipps, Antonia Körner, Jakob Vanhoefer, Dilan Pathirana, and Jan Hasenauer. Non- Negative Universal Differential Equations With Applications in Systems Biology. IF AC- PapersOnLine, 58(23):25–30, 2024. ISSN 24058963. doi: 10.1016/j.ifacol.2024.10.005. URL https://link...

    2024 doi

  17. [25]

    Python Language Reference, version 3.9

    Python Software Foundation. Python Language Reference, version 3.9 . Python Software Foundation, 2020. URL https://www.python.org/

    2020

  18. [26]

    Jeff Bezanson, Alan Edelman, Stefan Karpinski, and Viral B. Shah. Julia: A Fresh Approach to Numerical Computing. SIAM Review, 59(1):65–98, January 2017. ISSN 0036-1445, 1095-7200. doi: 10.1137/141000671. URL https://epubs.siam.org/doi/10.1137/141000671. 11

    2017 doi

  19. [27]

    Jordan, and Ion Stoica

    Philipp Moritz, Robert Nishihara, Stephanie Wang, Alexey Tumanov, Richard Liaw, Eric Liang, Melih Elibol, Zongheng Yang, William Paul, Michael I. Jordan, and Ion Stoica. Ray: A Distributed Framework for Emerging AI Applications, 2017. URL https://arxiv.org/ abs/1712.05889. Ver...

    2017 arXiv

  20. [2004]

    doi: 10.1002/cncr.11882

    ISSN 0008-543X, 1097-0142. doi: 10.1002/cncr.11882. URL https://acsjournals. onlinelibrary.wiley.com/doi/10.1002/cncr.11882

    doi

  21. [2023]

    doi: 10.1016/j.heliyon.2023.e17890

    ISSN 24058440. doi: 10.1016/j.heliyon.2023.e17890. URL https://linkinghub. elsevier.com/retrieve/pii/S2405844023050983

    2023 doi

  22. [7755]

    URL http://ascopubs.org/doi/10.1200/JCO

    doi: 10.1200/JCO.2002.02.140. URL http://ascopubs.org/doi/10.1200/JCO. 2002.02.140

    2002 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.