Clinical Plotting

Fyron's clinical plotting helpers are exposed through the dedicated fyron.plotting facade. They remain plain Matplotlib functions, so every figure can be saved, customized, combined into panels, or audited through returned tables.

Use these plots to explain clinical data science decisions: survival differences, model discrimination, calibration, thresholds, cohort flow, biomarker distributions, correlations, and manuscript-ready effect estimates.

Use This Page For

For CT/MR slice comparison figures, use BOA Visualization. Those helpers validate image geometry and return slice-level difference summary tables.

Install

uv add "fyron[ml]"
uv add "fyron[survival]"

Import

from fyron import plotting as fp

Existing imports continue to work:

from fyron import ml
from fyron.survival import plot_kaplan_meier

Core Concepts

API Contract

Plotting Philosophy

Clinical figures often need to be clear in slides, manuscripts, and grayscale printouts. Fyron therefore:

• uses Matplotlib, avoiding a large plotting stack,
• returns figure and axes objects for downstream customization,
• supports save_path on plotting functions,
• includes hatch patterns for grouped bar plots,
• defaults categorical plots to the colorblind-safe Okabe-Ito palette,
• keeps BOARISK/Bits & Flames colors available as named palettes such as bf, bf_vibrant, and bf_vibrant_muted,
• defaults correlation plots to Spearman for clinical exploratory work,
• includes grouped boxplots with patient-level jitter for cohort and hue comparisons,
• includes smoothed density plots for distributions by event, treatment, or risk group,
• includes grouped bar plots with statistical significance notation for manuscript summaries,
• includes Kaplan-Meier curves for clinical time-to-event endpoints,
• includes calibration and precision-recall plots, not only ROC curves.

Clinical Figure Preview

All previews below use synthetic data. They demonstrate figure layout, palette choice, and export style rather than clinical results.

Plot Selection Guide

Function Summary

Function Parameter Reference

Parameter Notes

Return Values

Most plotting functions return (fig, ax). Some functions return (fig, ax, table) when the computed table is part of the result, such as statistical bar plots or correlation heatmaps. Kaplan-Meier functions return a result object with plot handles and survival statistics. Keep returned tables with exported figures so visual annotations can be audited.

Scientific Palettes And Styles

Fyron uses named palettes so a study can keep the same visual language across notebooks, figures, slides, and supplements. The defaults follow scientific color guidance from SciFig, Datawrapper, and Molecular Ecologist: use qualitative colors for categories, sequential colors for ordered magnitude, and diverging colors when values have a meaningful midpoint. Avoid rainbow/jet maps for continuous data, keep categorical palettes small, and pair color with hatches, points, labels, or small multiples when a figure has many groups.

The default categorical palette is okabe_ito, a colorblind-safe palette widely used for scientific figures. BOARISK/Bits & Flames colors remain available as palette="bf", palette="bf_vibrant", and palette="bf_vibrant_muted" when brand continuity matters.

from fyron import plotting as fp

fp.list_palettes()

fig, ax = fp.preview_palette("okabe_ito")
audit = fp.palette_luminance_table("okabe_ito")
contrast = fp.palette_grayscale_contrast("okabe_ito")

fp.set_plot_style(style="paper", palette="okabe_ito")

Use named palettes and colormaps directly in plotting functions:

fig, ax = fp.plot_grouped_boxplot(
    clinical_features,
    x="risk_group",
    y="l3_sma_cm2",
    palette="okabe_ito",
)

km = fp.plot_kaplan_meier(
    cohort,
    duration_col="time",
    event_col="event",
    group_col="risk_group",
    palette="okabe_ito",
)

fig, ax, corr = fp.plot_correlation_heatmap(
    clinical_features,
    columns=["age", "tumor_volume", "risk_score"],
    cmap="fyron_diverging_blue_red",
)

Practical color rules:

• Use okabe_ito or wong for categorical clinical groups.
• Use muted_scientific or nature_cool for restrained journal-style categorical plots.
• Use bf_vibrant_muted for branded figures that should feel like Bits & Flames without overpowering manuscript layouts.
• Use fyron_sequential_blue or fyron_sequential_teal for ordered magnitude.
• Use fyron_diverging_blue_red or fyron_diverging_teal_orange for centered effects such as correlations, residuals, or signed changes.
• If a categorical palette needs more colors than recommended, Fyron warns. Consider hatches, marker shapes, faceting, or small multiples.
• Use external tools such as Viz Palette, Coblis, Sim Daltonism, or Datawrapper-style checks for full color-vision simulation before submission.

Hatch And Schraffierung Styles

Schraffierungen make grouped figures more robust in grayscale, print, slide decks, and color-vision-sensitive settings. Use them when color alone would make groups hard to distinguish, especially for bar plots with multiple cohorts, endpoints, models, or scanner groups. For swarmplots, density plots, and line plots, prefer points, markers, direct labels, or small multiples instead of hatch patterns.

Fyron uses these default hatch styles:

Use the hatches= parameter when you want to lock a specific visual language:

from fyron import plotting as fp

fig, ax = fp.plot_patterned_bar(
    summary_df,
    x="endpoint",
    y="rate",
    group="cohort",
    error="ci_half_width",
    palette="bf_vibrant_muted",
    hatches=["", "///", "\\\\\\"],
    annotate=True,
    title="Endpoint rates by cohort",
)

plot_metric_bars builds on the same patterned-bar style, so model-comparison bars inherit black outlines, grouped hatches, and grayscale-safe visual separation. For manuscript figures, combine hatches with restrained palettes, black outlines, direct labels, and exported vector formats such as PDF or SVG.

Statistical Annotations

Fyron includes a Matplotlib-native annotation layer inspired by trevismd/statannotations. It keeps the useful manuscript behavior, including pairwise tests, multiple-testing correction, custom p-values, star labels, explicit p-value labels, and bracket layout, without adding Seaborn as a required dependency.

Use stat_test_pairs when you need the statistics table separately, and use add_stat_annotations when you already have an axis to annotate:

stats = fp.stat_test_pairs(
    clinical_features,
    x="risk_group",
    y="risk_score",
    pairs=[("Low", "High")],
    test="mannwhitney",
    p_adjust="fdr_bh",
)

fp.add_stat_annotations(ax, stats)

For built-in categorical plots, prefer return_stats=True so the figure and its statistical audit table are produced together.

Model Evaluation Plot Set

from fyron import ml, plotting as fp

rf = ml.train_random_forest(
    X_train=X_train,
    y_train=y_train,
    X_test=X_test,
    y_test=y_test,
    n_estimators=300,
    random_state=42,
)

fig_roc, _ = fp.plot_roc_curve(
    y_true=y_test,
    y_prob=rf["y_prob"],
    title="Random Forest ROC",
)
fig_pr, _ = fp.plot_precision_recall_curve(
    y_true=y_test,
    y_prob=rf["y_prob"],
)
fig_cal, _ = fp.plot_calibration_curve(
    y_true=y_test,
    y_prob=rf["y_prob"],
    n_bins=10,
)
fig_cm, _ = fp.plot_confusion_matrix(
    y_true=y_test,
    y_pred=rf["y_pred"],
)

When plot=True is passed to train_random_forest or train_xgboost, these figures are created automatically:

rf = ml.train_random_forest(
    X_train=X_train,
    y_train=y_train,
    X_test=X_test,
    y_test=y_test,
    n_estimators=300,
    random_state=42,
    plot=True,
)

rf["figures"].keys()
# dict_keys(["roc", "precision_recall", "calibration", "confusion_matrix"])

For cross-validation, pass a table of out-of-fold predictions instead of a fitted model. This keeps model fitting separate from figure generation and makes the AUC table easy to save with the manuscript supplement.

cv_predictions = pd.DataFrame(
    {
        "fold": fold_ids,
        "y_true": y_true_oof,
        "y_prob": y_prob_oof,
    }
)

fig, ax, auc_table = fp.plot_cross_validated_roc(
    cv_predictions,
    title="Five-fold cross-validated ROC",
    palette="okabe_ito",
)

Correlation Heatmap

Use a correlation heatmap for feature screening, redundancy checks, and sanity checks before modeling. Fyron uses method="spearman" by default because clinical variables are often skewed, ordinal, or monotonic without being normally distributed.

fig, ax, corr = fp.plot_correlation_heatmap(
    data=clinical_features,
    columns=["age", "tumor_volume", "risk_score", "lab_value"],
    method="spearman",
    mask_upper=True,
    annotate_significance=True,
    p_adjust="fdr_bh",
    title="Feature correlation",
    cmap="fyron_diverging_teal_orange",
)

For subgroup analyses, use group_col or call plot_grouped_correlation_heatmaps directly:

fig, axes, corr_long = fp.plot_grouped_correlation_heatmaps(
    data=clinical_features,
    group_col="treatment_arm",
    columns=["age", "tumor_volume", "risk_score", "lab_value"],
    annotate_significance=True,
    cmap="fyron_diverging_teal_orange",
)

Tips:

• Use method="spearman" for ordinal or non-linear monotonic relationships.
• Use method="pearson" when a linear relationship is the actual assumption.
• Use mask_upper=True to reduce visual duplication in large matrices.
• The returned corr DataFrame can be saved as a supplement table.

Grouped Boxplots

Use grouped boxplots when the clinical question is about a continuous biomarker across discrete patient groups: L3 muscle area by risk group, creatinine by treatment arm, VAT by sex and cohort, or a model score by response category. Fyron draws each box with a black outline, restrained color fill, and optional patient-level jitter so outliers and sample size remain inspectable.

Required columns:

• x: categorical group column such as risk group, treatment arm, sex, stage, or endpoint.
• y: numeric marker column such as l3_sma_cm2, risk_score, vat_ml, or albumin.

Common optional parameters:

• hue: second grouping column drawn side by side inside each x group.
• order and hue_order: explicit display order for manuscript-ready figures.
• show_points: overlay patient-level jittered observations.
• palette: custom color sequence.

fig, ax = fp.plot_grouped_boxplot(
    data=clinical_features,
    x="risk_group",
    y="l3_sma_cm2",
    hue="sex",
    order=["Low", "Intermediate", "High"],
    hue_order=["Female", "Male"],
    title="L3 skeletal muscle area by risk group",
    ylabel="L3 SMA (cm2)",
    palette="okabe_ito",
)

Add statannotations-style brackets when the figure needs explicit comparisons. Fyron computes the tests with SciPy, applies Fyron p-value correction, and returns the audit table when return_stats=True.

fig, ax, stats = fp.plot_grouped_boxplot(
    data=clinical_features,
    x="risk_group",
    y="l3_sma_cm2",
    comparisons=[("Low", "High")],
    stat_test="mannwhitney",
    p_adjust="fdr_bh",
    annotation_format="star",
    return_stats=True,
    palette="okabe_ito",
)

For small and medium cohorts, keep show_points=True so readers can see how many observations drive each box. For very large cohorts, set show_points=False and consider a density plot next to the boxplot.

Statistical Grouped Bar Plots

Use statistical grouped bar plots when the manuscript needs compact mean or median summaries with explicit significance notation. Fyron draws grouped/hue bars with black outlines, optional error bars, and bracket annotations. The returned table includes one row per plotted summary plus one row per statistical comparison, so the visual notation can be audited.

Defaults:

• estimator="mean" and error="sem".
• test="auto" uses Mann-Whitney U for pairwise group comparisons.
• hue with two levels compares hue levels inside each x group.
• two x groups without hue compare the two x groups directly.
• notation defaults to ***, **, *, and ns.

fig, ax, stats_table = fp.plot_grouped_bar_with_significance(
    data=clinical_features,
    x="risk_group",
    y="l3_sma_cm2",
    hue="sex",
    estimator="mean",
    error="sem",
    title="L3 skeletal muscle area by risk group and sex",
    ylabel="L3 SMA (cm2)",
)

For small cohorts, pair this plot with a grouped boxplot or density plot so readers can inspect the distribution behind the summary bars.

Density Plots

Density plots are helpful when a boxplot hides the actual shape of the distribution. Use them to compare risk scores between event and non-event patients, lab values between treatment arms, or imaging markers between anatomical subgroups. Fyron uses a lightweight NumPy Gaussian kernel density estimate, so no Seaborn or SciPy dependency is required.

Required columns:

• x: numeric marker or score column.

Common optional parameters:

• hue: categorical column for separate curves.
• hue_order: explicit curve order.
• fill: fill the area below each curve.
• rug: show patient-level tick marks along the baseline.
• bandwidth: custom smoothing bandwidth when the automatic rule is too smooth or too jagged.
• gridsize: number of x-axis evaluation points.

fig, ax = fp.plot_density(
    data=clinical_features,
    x="risk_score",
    hue="event",
    hue_order=[0, 1],
    title="Risk score density by event status",
    xlabel="Predicted risk score",
    rug=True,
    palette="okabe_ito",
)

Interpret density plots as distribution summaries, not counts. If absolute group size matters in the figure, use plot_histogram; if each patient-level value should remain visible, use plot_swarmplot.

fig, ax = fp.plot_histogram(
    clinical_features,
    x="risk_score",
    hue="event",
    stat="percent",
    multiple="dodge",
    palette="okabe_ito",
)

fig, ax, stats = fp.plot_swarmplot(
    clinical_features,
    x="risk_group",
    y="risk_score",
    hue="sex",
    comparisons=[(("High", "Female"), ("High", "Male"))],
    return_stats=True,
    palette="okabe_ito",
)

Validation And QC Plots

These plots consume tables produced by fyron.validation, fyron.reporting, or fyron.imaging.segmentation.

from fyron import plotting as fp

fig_cal, _ = fp.plot_survival_calibration(calibration_bins)
fig_dca, _ = fp.plot_decision_curve(decision_curve)
fig_flow, _ = fp.plot_cohort_flow(flow_table)
fig_dice, _ = fp.plot_dice_by_label(segmentation_metrics)
fig_ext, _ = fp.plot_external_validation_metrics(
    {"auc": 0.81, "sensitivity": 0.74, "specificity": 0.69}
)

When inclusion and exclusion filters are executed in code, generate the count table with fyron.flow and pass it directly to plot_cohort_flow:

from fyron.flow import FlowTracker
from fyron import plotting as fp

flow = FlowTracker("NSCLC CT cohort")
flow.add_step("Source cohort", raw)
flow.add_step("Has baseline CT", with_ct, reason="missing CT")

fig, ax = fp.plot_cohort_flow(flow.to_dataframe())

For segmentation visual QC:

fig, axes = fp.plot_segmentation_overlay_grid(
    image=ct_array,
    mask=label_map,
    slice_indices=[40, 60, 80],
    label=5,
)

Kaplan-Meier Curves

Kaplan-Meier curves belong in the clinical plotting toolkit because many cohort studies, imaging biomarkers, and risk scores are evaluated against time-to-event endpoints. You can call them through fyron.plotting.plot_kaplan_meier with the other clinical plots; internally the function delegates to fyron.survival because it fits Kaplan-Meier estimators, computes log-rank comparisons, and can annotate restricted mean survival time.

Required columns:

• duration_col: positive follow-up time.
• event_col: event indicator where 1 means event and 0 means censored.

Common optional columns and parameters:

• group_col: categorical group such as stage, treatment, or risk group.
• split_column: continuous score to split into median, tertile, quartile, or custom groups.
• at_risk_counts: add an at-risk count table below the plot.
• show_censors: show censoring ticks.
• xlim: limit the time axis for a prespecified follow-up horizon.
• annotate_logrank: add the log-rank p-value to the plot.
• annotate_rmst: annotate restricted mean survival time when configured.

from fyron import plotting as fp

km = fp.plot_kaplan_meier(
    df=cohort,
    duration_col="time",
    event_col="event",
    group_col="risk_group",
    at_risk_counts=True,
    show_censors=True,
    annotate_logrank=True,
    xlim=(0, 60),
    title="Overall survival by risk group",
    figsize=(7, 4),
    palette="okabe_ito",
)

km.figure.savefig("figures/km_risk_group.png", bbox_inches="tight")
km.logrank
km.n_per_group

Use split_column when you have a continuous score and want Fyron to create the groups before plotting:

km = fp.plot_kaplan_meier(
    df=cohort,
    duration_col="time",
    event_col="event",
    split_column="risk_score",
    split_method="quartile",
    at_risk_counts=True,
    title="Overall survival by risk-score quartile",
    palette="okabe_ito",
)

For full survival modeling details, see Survival Analysis.

Plot Families

The BOARISK papers use a compact visual language: light grids, black outlines, restrained colors, and hatches so grouped figures remain readable in grayscale. Fyron exposes the same language in general-purpose functions for clinical data science.

Risk Score Distribution

Use this when you have a risk score, model score, or survival risk output and want to see whether event-positive patients are shifted toward higher values. It accepts a plain table and can draw one combined distribution or separate groups by event_col.

Required columns:

• risk_col: numeric score column.

Optional columns and parameters:

• event_col: grouping column, usually 0 for censored/non-event and 1 for event.
• event_labels: mapping or sequence used to replace group labels.
• jitter: overlay patient-level points.
• violin: use violin distributions; set to False for box plots.

import pandas as pd
from fyron import plotting as fp

risk_df = pd.DataFrame(
    {
        "risk_score": [0.12, 0.24, 0.35, 0.61, 0.72, 0.89],
        "event": [0, 0, 0, 1, 1, 1],
    }
)

fig, ax = fp.plot_risk_score_distribution(
    data=risk_df,
    risk_col="risk_score",
    event_col="event",
    event_labels={0: "Censored", 1: "Event"},
    title="Risk score by event status",
    palette="bf_vibrant_muted",
)

Signed Feature Weights

Use this for Cox coefficients, logistic regression coefficients, SHAP-like signed summaries, or any model output where positive values push risk upward and negative values push risk downward.

Required columns:

• feature_col: feature or covariate name.
• weight_col: signed numeric weight.

Optional parameters:

• top_n: keep the largest absolute weights.
• positive_label and negative_label: legend labels for interpretation.

coef_df = pd.DataFrame(
    {
        "feature": ["L3 SMA", "VAT", "Age", "Albumin", "Tumor volume"],
        "coef": [-0.42, 0.38, 0.29, -0.31, 0.67],
    }
)

fig, ax = fp.plot_signed_feature_weights(
    data=coef_df,
    feature_col="feature",
    weight_col="coef",
    top_n=5,
    title="Top signed Cox coefficients",
)

Significance Bars

Use this for log-rank results, subgroup tests, univariate screening, or endpoint comparisons. The plot converts p-values to -log10(p) so smaller p-values are visually stronger. It also returns the computed table with minus_log10_p.

Required columns:

• label_col: endpoint, subgroup, feature, or comparison name.
• p_col: p-value column.

Optional parameters:

• threshold: significance threshold drawn as a dashed line.
• annotate: show readable p-value labels next to bars.

p_values = pd.DataFrame(
    {
        "comparison": ["L3 SMA", "VAT", "Albumin", "Age"],
        "p_value": [0.003, 0.048, 0.021, 0.19],
    }
)

fig, ax, table = fp.plot_significance_bars(
    data=p_values,
    label_col="comparison",
    p_col="p_value",
    threshold=0.05,
    title="Log-rank screening",
)

Directional Ratio Rows

Use this for odds ratios, hazard ratios, or risk ratios when the direction matters as much as the estimate. Ratios below 1.0 are shown on the protective side, ratios above 1.0 on the hazardous side, and each feature receives one row.

Required columns:

• feature_col: feature, biomarker, subgroup, or covariate name.
• ratio_col: numeric odds ratio, hazard ratio, or risk ratio.

Optional columns:

• lower_col and upper_col: confidence interval bounds.
• p_col: p-value annotation.
• group_col: category prefix for row labels.
• n_col: sample size appended to row labels.

The right-side information table is optional. Keep annotate=True to show estimate, 95% CI, and p value columns; set annotate=False for a pure visual ratio plot. Fyron infers HR, OR, or RR from ratio_label, and you can override the headers with estimate_label, ci_label, and p_label.

ratio_df = pd.DataFrame(
    {
        "feature": ["Sarcopenia", "High VAT", "Albumin", "Stage IV"],
        "category": ["Body composition", "Body composition", "Laboratory", "Clinical"],
        "hazard_ratio": [1.78, 1.41, 0.68, 2.16],
        "ci_lower": [1.18, 1.03, 0.49, 1.44],
        "ci_upper": [2.68, 1.94, 0.94, 3.24],
        "p_value": [0.006, 0.031, 0.018, 0.0004],
    }
)

fig, ax = fp.plot_directional_ratio_rows(
    data=ratio_df,
    feature_col="feature",
    ratio_col="hazard_ratio",
    lower_col="ci_lower",
    upper_col="ci_upper",
    p_col="p_value",
    group_col="category",
    ratio_label="Hazard ratio",
    estimate_label="HR",
    ci_label="95% CI",
    p_label="p value",
    title="Feature-level hazard ratios",
)

Forest Estimates

Use this for Cox hazard ratios, odds ratios, risk ratios, subgroup analyses, or any estimate with lower and upper confidence bounds. Fyron defaults to a log x-axis because ratios are usually easier to read that way.

Required columns:

• label_col: row label.
• estimate_col: point estimate.
• lower_col: lower confidence interval.
• upper_col: upper confidence interval.

Optional columns:

• p_col: p-value annotation.
• p_adjust_col: adjusted p-value column when you want FDR-aware labels.
• n_col: sample size appended to the row label.
• table_columns: append estimate, CI, and p-value text to the y-axis labels.

forest_df = pd.DataFrame(
    {
        "subgroup": ["Low sarcopenia", "Intermediate", "High sarcopenia"],
        "hazard_ratio": [0.82, 1.0, 1.74],
        "ci_lower": [0.52, 0.73, 1.15],
        "ci_upper": [1.29, 1.37, 2.61],
        "p_value": [0.37, 0.99, 0.008],
        "n": [118, 142, 96],
    }
)

fig, ax = fp.plot_forest_estimates(
    data=forest_df,
    label_col="subgroup",
    estimate_col="hazard_ratio",
    lower_col="ci_lower",
    upper_col="ci_upper",
    p_col="p_value",
    n_col="n",
    title="Overall survival subgroup effects",
    xlabel="Hazard ratio",
    table_columns=True,
)

For a Fyron Cox result, use the Cox-specific convenience wrapper:

cox = fit_multivariate_cox(
    cohort,
    duration_col="time",
    event_col="event",
    covariates=["age", "stage", "risk_score"],
)

fig, ax = fp.plot_hazard_ratios(cox, title="Cox model hazard ratios")

Likert Scale Plots

Use plot_likert_scale for reader studies, usability surveys, annotation review, or clinical expert ratings. The input should be a long table with one row per response, or one row per response category with a count column.

fig, ax, likert_table = fp.plot_likert_scale(
    data=reader_survey,
    item_col="item",
    response_col="response",
    count_col="count",
    response_order=[
        "Disagree",
        "Neutral",
        "Agree",
        "Strongly agree",
    ],
    neutral_values=["Neutral"],
    palette="okabe_ito",
    title="Reader study agreement",
)

The returned table contains count, percent, side, and plot_value, so the plotted values can be exported with the figure.

Waffle Plots

Use plot_waffle for simple cohort composition, class balance, endpoint mix, or review status summaries. Waffle plots are strongest when the categories are few and the total has an intuitive interpretation, such as 100 squares.

fig, ax, waffle_table = fp.plot_waffle(
    data=cohort_mix,
    category_col="category",
    value_col="n",
    total=100,
    rows=10,
    palette="bf_vibrant_muted",
    title="Cohort composition",
)

The returned table includes source values, proportions, and allocated cell counts. For precise reporting, use the table values in text and the waffle plot as a visual summary.

Patterned Bar Plot

Patterned bars are useful when groups need to remain distinguishable without color.

fig, ax = fp.plot_patterned_bar(
    data=summary_df,
    x="endpoint",
    y="rate",
    group="cohort",
    error="ci_half_width",
    palette="bf_vibrant_muted",
    hatches=["", "///"],
    title="Endpoint rates by cohort",
    ylabel="Rate",
    annotate=True,
)

Expected input:

Model Metric Bars

plot_metric_bars accepts either a nested dictionary or a DataFrame.

fig, ax = fp.plot_metric_bars(
    metrics={
        "Random Forest": rf["metrics"],
        "XGBoost": xgb["metrics"],
    },
    metrics_to_plot=["auc", "sensitivity", "specificity", "f1"],
    title="Model comparison",
)

This is a compact way to compare clinically meaningful metrics instead of focusing only on accuracy.

Saving Figures

Every plotting helper accepts save_path.

fp.plot_calibration_curve(
    y_true=y_test,
    y_prob=rf["y_prob"],
    save_path="figures/calibration_rf.png",
)

You can also use regular Matplotlib calls:

fig, ax = fp.plot_roc_curve(
    y_true=y_test,
    y_prob=rf["y_prob"],
)
ax.set_title("Internal validation ROC")
fig.savefig("figures/roc_internal.svg", bbox_inches="tight")

Choosing The Right Plot