Many thanks to all our co-authors on the paper for all their hard work and support!
8/8

The Periskope-CM models accurately predict mortality risk in patients with HIV-associated CM and show promise as a tool to direct future treatment stratification trials. They are available to use for research purposes at the following link: www.periskope.org/cm/
7/8

Explorative analysis of treatment effect demonstrated a trend towards improved outcomes in patients with low predicted risk who received the single high dose liposomal amphotericin B, compared with standard amphotericin B deoxycholate + flucytosine.
6/8

Calibration, the measure of how well model predictions matched observed events, was reasonable for both regression models, although demonstrated some overprediction of risk at high-risk ranges, where training data was sparser. The machine learning model demonstrated miscalibration.
5/8

Both regression models discriminated well between patients who met the outcome and those who did not in held out validation data. The machine learning model did not improve performance.
4/8

We developed multivariable logistic regression models to predict 2-week mortality on data from the ACTA www.nejm.org/doi/full/10.... and Ambition www.nejm.org/doi/full/10.... trials using best-practice methods, comparing them head-to-head with a machine learning model trained on the same data.
3/8

Cryptococcal meningitis (CM) is a major driver of global HIV-related deaths. Treatment is intensive and toxic, so stratifying patients by disease severity may allow for more targeted therapy – less intensive regimens for less unwell patients.
2/8

Many thanks to all our co-authors on the paper for all their hard work and support!
8/8

The Periskope-CM models accurately predict mortality risk in patients with HIV-associated CM and show promise as a tool to direct future treatment stratification trials. They are available to use for research purposes at the following link: www.periskope.org/cm/
7/8

Explorative analysis of treatment effect demonstrated a trend towards improved outcomes in patients with low predicted risk who received the single high dose liposomal amphotericin B, compared with standard amphotericin B deoxycholate + flucytosine.
6/8

Calibration, the measure of how well model predictions matched observed events, was reasonable for both regression models, although demonstrated some overprediction of risk at high-risk ranges, where training data was sparser. The machine learning model demonstrated miscalibration.
5/8

Both regression models discriminated well between patients who met the outcome and those who did not in held out validation data. The machine learning model did not improve performance.
4/8

We developed multivariable logistic regression models to predict 2-week mortality on data from the ACTA www.nejm.org/doi/full/10.... and Ambition www.nejm.org/doi/full/10.... trials using best-practice methods, comparing them head-to-head with a machine learning model trained on the same data.
3/8

Cryptococcal meningitis (CM) is a major driver of global HIV-related deaths. Treatment is intensive and toxic, so stratifying patients by disease severity may allow for more targeted therapy – less intensive regimens for less unwell patients.
2/8