CoChat

Deep learning for small and big data in psychiatry

Open access
Koppe, Georgia, Meyer-Lindenberg, Andreas, Durstewitz, Daniel · 2021 · Neuropsychopharmacology

The DOI matches its Crossref record; 170 of the 170 references it cites were found in a registry or an index.

Abstract

Psychiatry today must gain a better understanding of the common and distinct pathophysiological mechanisms underlying psychiatric disorders in order to deliver more effective, person-tailored treatments. To this end, it appears that the analysis of ‘small’ experimental samples using conventional statistical approaches has largely failed to capture the heterogeneity underlying psychiatric phenotypes. Modern algorithms and approaches from machine learning, particularly deep learning, provide new hope to address these issues given their outstanding prediction performance in other disciplines. The strength of deep learning algorithms is that they can implement very complicated, and in principle arbitrary predictor-response mappings efficiently. This power comes at a cost, the need for large training (and test) samples to infer the (sometimes over millions of) model parameters. This appears to be at odds with the as yet rather ‘small’ samples available in psychiatric human research to date (n < 10,000), and the ambition of predicting treatment at the single subject level (n = 1). Here, we aim at giving a comprehensive overview on how we can yet use such models for prediction in psychiatry. We review how machine learning approaches compare to more traditional statistical hypothesis-driven approaches, how their complexity relates to the need of large sample sizes, and what we can do to optimally use these powerful techniques in psychiatric neuroscience.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41386-020-0767-z

Cites

170 of 170 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Statistical hypothesis testing vs. prediction in machine learning. The philosophy in classical statistical hypothesis testing (bottom path) is to draw a random sample from a population and estimate parameters of a model, which is assumed to describe the population sufficiently well (hence does not need to be selected among a larger class of models). Hypotheses about the population are then tested in terms of the model parameters. For instance, one may test whether there is a linear relationship between a feature and an output by formulating a null hypothesis on the slope parameter β1. In prediction, in contrast, which is what most machine learning methods aim for, we should look for the model which is the best in predicting outcomes in new samples (purple dots). Hence, rather than settling on one model a priori which is believed to describe the statistical properties of the true population, multiple models are trained in order to select the one which minimizes the loss on an independent validation set (blue dots).
Fig. 1 Statistical hypothesis testing vs. prediction in machine learning. The philosophy in classical statistical hypothesis testing (bottom path) is to draw a random sample from a population and estimate parameters of a model, which is assumed to describe the population sufficiently well (hence does not need to be selected among a larger class of models). Hypotheses about the population are then tested in terms of the model parameters. For instance, one may test whether there is a linear relationship between a feature and an output by formulating a null hypothesis on the slope parameter β1. In prediction, in contrast, which is what most machine learning methods aim for, we should look for the model which is the best in predicting outcomes in new samples (purple dots). Hence, rather than settling on one model a priori which is believed to describe the statistical properties of the true population, multiple models are trained in order to select the one which minimizes the loss on an independent validation set (blue dots).CC BY · PubMed Central
Fig. 2 Feed-forward neural networks (FNNs) and function approximation. a Schematic of a logistic regression model (top) and FNN model (bottom) to predict an output (here symptom probability) from two types of features (e.g., brain function and structure). While the logistic regression model directly maps the weighted inputs through a logistic (sigmoid) type function, the FNN first filters the weighted inputs in successive stages by propagating them through multiple layers of units, with a nonlinear, e.g., sigmoid, activation function. While the logistic regression model can only separate two features linearly, the recombination of inputs across multiple stages allows the FNN to implement quite complex (in fact, arbitrarily complex) input output mappings (right panel). b In other words, the FNN has a much larger space of functions it can implement and thus a higher model complexity, including logistic regression functions as special cases, and may therefore be able to infer a function (denoted by f^FNN) closer to the true function f. c FNNs can be constructed with different types of activation functions such as ReLUs, sigmoids, or radial basis functions (RBFs; top panel). A strength of ReLUs is that their derivative is piecewise constant, whereas sigmoids and RBFs may have strongly varying gradients and saturate at the extremes (bottom panel).
Fig. 2 Feed-forward neural networks (FNNs) and function approximation. a Schematic of a logistic regression model (top) and FNN model (bottom) to predict an output (here symptom probability) from two types of features (e.g., brain function and structure). While the logistic regression model directly maps the weighted inputs through a logistic (sigmoid) type function, the FNN first filters the weighted inputs in successive stages by propagating them through multiple layers of units, with a nonlinear, e.g., sigmoid, activation function. While the logistic regression model can only separate two features linearly, the recombination of inputs across multiple stages allows the FNN to implement quite complex (in fact, arbitrarily complex) input output mappings (right panel). b In other words, the FNN has a much larger space of functions it can implement and thus a higher model complexity, including logistic regression functions as special cases, and may therefore be able to infer a function (denoted by f^FNN) closer to the true function f. c FNNs can be constructed with different types of activation functions such as ReLUs, sigmoids, or radial basis functions (RBFs; top panel). A strength of ReLUs is that their derivative is piecewise constant, whereas sigmoids and RBFs may have strongly varying gradients and saturate at the extremes (bottom panel).CC BY · PubMed Central
Fig. 3 Different types of prediction errors. a Contours of two Gaussian distributions associated with two fictional populations (red and blue) showing a probabilistic relationship between a feature and an outcome (e.g., brain volume reduction and age). The ellipses mark points of equal probability density at standard deviations σ = 1, 2, 3, indicating the spread of the Gaussians. The red population shows slightly less spread (potentially related to stricter inclusion criteria or differences in the measurement device used for this population). b Two random samples of n = 30 points drawn from both distributions (indicated by corresponding colors). c 50% of the red sample (depicted in B) is used to fit a linear model (thick skewed red line). The remaining 50% of sample points (the test set), here displayed as white circles, are used to evaluate the out-of-sample error (red vertical lines). Another sample of outcomes is drawn at the exact same feature values used for training (orange circles) and used to evaluate the in-sample prediction error (orange vertical lines). d The model (red line, same as in c) is now employed to predict the outcome for the blue (more broad) sample (potentially collected at a different site). The blue vertical lines mark the out-of-domain prediction error. This error appears to be larger than both the other errors (c) and indicates a systematic underestimation of the outcome.
Fig. 3 Different types of prediction errors. a Contours of two Gaussian distributions associated with two fictional populations (red and blue) showing a probabilistic relationship between a feature and an outcome (e.g., brain volume reduction and age). The ellipses mark points of equal probability density at standard deviations σ = 1, 2, 3, indicating the spread of the Gaussians. The red population shows slightly less spread (potentially related to stricter inclusion criteria or differences in the measurement device used for this population). b Two random samples of n = 30 points drawn from both distributions (indicated by corresponding colors). c 50% of the red sample (depicted in B) is used to fit a linear model (thick skewed red line). The remaining 50% of sample points (the test set), here displayed as white circles, are used to evaluate the out-of-sample error (red vertical lines). Another sample of outcomes is drawn at the exact same feature values used for training (orange circles) and used to evaluate the in-sample prediction error (orange vertical lines). d The model (red line, same as in c) is now employed to predict the outcome for the blue (more broad) sample (potentially collected at a different site). The blue vertical lines mark the out-of-domain prediction error. This error appears to be larger than both the other errors (c) and indicates a systematic underestimation of the outcome.CC BY · PubMed Central
Fig. 4 Model complexity and the bias-variance trade-off. a As model complexity is increased (x-axis), variance rises and bias declines, that is, lower bias is traded for higher variance. We want to select the (optimal) model which balances these two quantities, achieving minimum prediction error (y-axis, minimum of bias plus variance, black curve). Increasing sample size effectively shifts this minimum to the right (dotted lines), enabling models of higher complexity. b Illustration of underfitting (top) and overfitting (bottom). Both panels depict the same samples (gray dots) drawn with noise from the true function (gray). The low and high complexity linear regression models with polynomial basis expansion of order 1 (top) and 20 (bottom) were fit (thick black lines). Panels depict the model deviation to the true function (blue lines) illustrating model bias. c Overfitting in detail: Here we assume that the true relation between inputs and outputs is perfectly linear as depicted by the black line (with 5 data points on that line illustrated). Assuming we have only observed one data point (black solid circle), we can however fit infinitely many lines (some of them illustrated in color) equally well. In this simple example, increasing the sample size just by one data point (and assuming there is no noise in the data) will allow us to pick out the correct model.
Fig. 4 Model complexity and the bias-variance trade-off. a As model complexity is increased (x-axis), variance rises and bias declines, that is, lower bias is traded for higher variance. We want to select the (optimal) model which balances these two quantities, achieving minimum prediction error (y-axis, minimum of bias plus variance, black curve). Increasing sample size effectively shifts this minimum to the right (dotted lines), enabling models of higher complexity. b Illustration of underfitting (top) and overfitting (bottom). Both panels depict the same samples (gray dots) drawn with noise from the true function (gray). The low and high complexity linear regression models with polynomial basis expansion of order 1 (top) and 20 (bottom) were fit (thick black lines). Panels depict the model deviation to the true function (blue lines) illustrating model bias. c Overfitting in detail: Here we assume that the true relation between inputs and outputs is perfectly linear as depicted by the black line (with 5 data points on that line illustrated). Assuming we have only observed one data point (black solid circle), we can however fit infinitely many lines (some of them illustrated in color) equally well. In this simple example, increasing the sample size just by one data point (and assuming there is no noise in the data) will allow us to pick out the correct model.CC BY · PubMed Central
Fig. 5 DNNs for individualized (treatment) predictions. In order to employ more complex FNN or RNN models for person-tailored predictions, we can pre-train a NN on multiple individuals. We first reduce the input dimensions, e.g., with an autoencoder (step 1), and then pre-train a DNN on the reduced inputs for a large sample (step 2). The pre-trained network may then be fine-tuned on the specific individual in a third step (right panel). Future data points could then be used to forecast symptom onset, treatment response, or other mental health-related variables.
Fig. 5 DNNs for individualized (treatment) predictions. In order to employ more complex FNN or RNN models for person-tailored predictions, we can pre-train a NN on multiple individuals. We first reduce the input dimensions, e.g., with an autoencoder (step 1), and then pre-train a DNN on the reduced inputs for a large sample (step 2). The pre-trained network may then be fine-tuned on the specific individual in a third step (right panel). Future data points could then be used to forecast symptom onset, treatment response, or other mental health-related variables.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 170 references

Research works printed but not linked: 13

  1. Montufar GF, Pascanu R, Cho K, Bengio Y. On the number of linear regions of deep neural networks. Advances in neural information processing systems. Curran Associates, Inc.; 2014. p. 2924–32.
  2. Ma X, Yang H, Chen Q, Huang D, Wang Y. Proceedings of the 6th International Workshop on Audio/Visual Emotion Challenge; 2016. p. 35–42.
  3. Yang T, Wu C, Huang K, Su M. in 2016 10th International Symposium on Chinese Spoken Language Processing (ISCSLP). IEEE; 2016. p. 1–5.
  4. Jarrett K, Kavukcuoglu K, Ranzato MA, LeCun Y. 2009 IEEE 12th international conference on computer vision. IEEE; 2009. p. 2146–53.
  5. Bishop CM. Pattern recognition and machine learning. New York, USA: Springer; 2006.
  6. Ruder S An overview of gradient descent optimization algorithms. arXiv preprint arXiv:160904747. 2016.
  7. Szegedy C, Zaremba W, Sutskever I, Bruna J, Erhan D, Goodfellow I, et al. Intriguing properties of neural networks. arXiv preprint arXiv:13126199. 2013.
  8. Ballard DH in AAAI 279-84 (1987).
  9. Gupta A, Ayhan M, Maida A in International Conference on Machine Learning; 2013. p. 987–94.
  10. Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, et al. Advances in neural information processing systems; 2014. p. 2672–80.
  11. Bowles C, Chen L, Guerrero R, Bentley P, Gunn R, Hammers A, et al. Gan augmentation: Augmenting training data using generative adversarial networks. arXiv preprint arXiv:181010863. 2018.
  12. Lee D-H. The simple and efficient semi-supervised learning method for deep neural networks. Workshop on challenges in representation learning, ICML. 2013:3.
  13. Yosinski J, Clune J, Bengio Y, Lipson H. How transferable are features in deep neural networks? Advances in neural information processing systems. Montreal, Canada: MIT Press; 2014. p 3320–28.

Cited by

See all 207

Metadata

Authors
Koppe, G., Meyer-Lindenberg, A., & Durstewitz, D.
Affiliations
Heidelberg University (DE)
Corresponding author
  • Koppe, Georgia
  • Durstewitz, Daniel
Venue
Neuropsychopharmacology
Published
2020-07-15
Volume
46
Issue
1
Pages
176-190
ISSN
  • 0893-133X (ISSN-L)
  • 1740-634X
OpenAlex
W3042578835
Cited by
189
References
168
Citations by year
2026: 16 · 2025: 27 · 2024: 55 · 2023: 34 · 2022: 36 · 2021: 17 · 2020: 4
FWCI
0.78
Type
journal article
Language
en
Open access
Licence
Locations
  • Neuropsychopharmacology (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Informationszentrum Lebenswissenschaften (ZB MED) (submittedVersion, cc-by)
Funders
  • Deutsche Forschungsgemeinschaft (CRC TRR 265: Project A06, CRC TRR 265: Project B08, ME 1591/4-1, 5485966, CRC636: Project B7, 255156212, CRC1158: Project B09, GRK 2350/1 Research Training Group 2350: Project B2, CRC TRR 265: Project S02, Du 354/8-2, Du 354/10-1)
  • Bundesministerium für Bildung und Forschung (01EF1803A, 01ZX1904A)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Functional Brain Connectivity Studies
Keywords
deep learning, machine learning, mental illness, small sample size, treatment prediction, personalized treatment, psychiatric phenotypes, sample size requirements, model complexity, psychiatric neuroscience
MeSH terms
Machine Learning, Deep Learning, Big Data, Mental Disorders, Humans, Psychiatry

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Deep learning for small and big data in psychiatry. https://app.cochat.ai/papers/10.1038/s41386-020-0767-z
The paper
Koppe, G., Meyer-Lindenberg, A., & Durstewitz, D. (2021). Deep learning for small and big data in psychiatry. Neuropsychopharmacology. https://doi.org/10.1038/s41386-020-0767-z

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.