DANDI:000469

In total, 41 recording sessions were completed across 21 patients with intractable localization-related epilepsy (Table 1 ). ... Screening sessions are available for all subjects except for 'P47HMH', for whom screening was not performed.

Methods, Subjects, paragraph 1; Table 1 and its legend

The files hold 21 subject IDs, 41 sessions (20 screening, 21 Sternberg) and one subject without a screening session, subject 19 (P47HMH). Subjects 13 and 14 have the native IDs P39CS and P39CS_2, age 26, male, recorded in 2015, with different units and electrodes (subject 13 has 48 and 33 electrodes in its two files, subject 14 has 34 and 34, at different locations); the files do not say whether they are one person recorded twice, so the 21 subjects may be 20 patients. Ages in the files run from 17 to 70, as in Table 1; the text gives 17 to 71.

In the Sternberg task, we recorded a total of 902 neurons (190 hippocampus, 259 Amygdala, 171 dorsal anterior cingulate cortex, 250 pre-supplementary motor area, 32 ventromedial prefrontal cortex) (Fig. 3b ).

Technical Validation, Data recorded and spike sorting quality metrics, paragraph 1; Figure 3b

Every count equals the paper's (902 units; 190 hippocampus, 259 amygdala, 171 dACC, 250 pre-SMA, 32 vmPFC). The notebook gives the counts per subject.

In the screening task, we recorded 907 neurons (Fig. 3a ). Of these, 449 were recorded in the MTL.

Technical Validation, Data recorded and spike sorting quality metrics, paragraph 1; Figure 3a

907 screening units, of which 421 in the MTL (159 hippocampus, 262 amygdala) and 486 in the MFC. 421 is within 10% of both readings of the criterion, 449 and 421, and equals the 421 implied by the paper's screening percentages (138 / 0.3278). 449 is the Sternberg MTL count (190 + 259), so the text gives the Sternberg number for the screening task.

During the working memory task (Fig. 1b ), subjects answered on average 88.73% of trials correctly (Fig. 1d ).

Technical Validation, Behavior, paragraph 1; Figure 1d

Mean over the 21 sessions 88.73% (s.e.m. 1.25, sessions 75.00 to 95.56%).

Accuracy and reaction time varied as a function of load as expected (Fig. 1c ), with higher loads resulting in slower and less accurate answers (F 2,40 = 7.92, P = 0.0013, repeated-measures ANOVA of median reaction times).

Technical Validation, Behavior, paragraph 1; Figure 1c

What the paper reports

• repeated-measures ANOVA of median reaction times, F(2,40) = 7.92, p = 0.0013 (n = 21 Sternberg sessions (df 2 and 40))

On correct trials the mean of the session medians is 1.055, 1.130 and 1.198 s for loads 1 to 3, and the load effect of the additive session by load ANOVA is F(2,40) = 7.92, p = 0.0013, the paper's values. On all trials (variant) F(2,40) = 4.08, p = 0.024. 17 of 21 sessions are slower at load 3 than at load 1.

How the test differed from the paper’s method

• The repeated-measures ANOVA is computed as the code's anovan model (session random, load fixed, no interaction), whose F test of load is the additive two-way ANOVA's; the correct-trial filter is the code's, as the criterion says.

Accuracy and reaction time varied as a function of load as expected (Fig. 1c ), with higher loads resulting in slower and less accurate answers

Technical Validation, Behavior, paragraph 1; Figure 1c

Mean accuracy over sessions 91.22% (load 1), 88.62% (load 2) and 86.35% (load 3), falling monotonically. The notebook gives the per-subject values; the paper reports no test.

On average across all neurons in the Sternberg task, ( i ) refractory period violations were 0.41% ± 0.50% of interspike intervals (ISIs) (Fig. 3d )

Technical Validation, Data recorded and spike sorting quality metrics, paragraph 2; Figure 3d

Over all 902 units (main reading) the mean is 0.428% (within tolerance) and the s.d. 0.625% (outside 0.50 plus or minus 0.05). With the code's filter (units below 4%), which leaves out one unit at 11.74%, the values are 0.416% and 0.498%, both within tolerance. The claim holds on the code's reading only; the measured values are that reading's.

How the test differed from the paper’s method

• The match needs the code's exclusion of units with 4% or more violations, which the paper does not state; it removes one unit.

(ii) average firing rate was 2.55 Hz, with a range of 0.11–33.82 Hz (Fig. 3e )

Technical Validation, Data recorded and spike sorting quality metrics, paragraph 2; Figure 3e

Mean 2.546 Hz, minimum 0.108 Hz, maximum 33.824 Hz over 902 units (median 1.387 Hz).

( iii ) the pairwise projection distance in clustering space between neurons isolated on the same wire was 13.77 ± 7.32 (projection test; in units of s.d. of the signal)

Technical Validation, Data recorded and spike sorting quality metrics, paragraph 2; Figure 3i

734 of 902 units have a non-zero stored projection distance; their mean is 13.774 and s.d. 7.325 (range 3.72 to 54.13).

( iv ) the ratio between the s.d. of noise and the peak mean waveform amplitude was 6.62 ± 3.63 (Fig. 3g,h )

Technical Validation, Data recorded and spike sorting quality metrics, paragraph 2; Figure 3g,h

Reading (a), the stored peak SNR, gives 6.625 plus or minus 3.632, within tolerance. Reading (b), the stored mean SNR, gives 1.647 plus or minus 0.954 and does not match.

(v) the modified coefficient of variation of the ISIs (CV2 52 ) was 0.94 ± 0.14 (Fig. 3f )

Technical Validation, Data recorded and spike sorting quality metrics, paragraph 2; Figure 3f

CV2 0.938 plus or minus 0.138 over 902 units, none undefined.

(vi) median isolation distance 53 was 31.14 (Fig. 3j )

Technical Validation, Data recorded and spike sorting quality metrics, paragraph 2; Figure 3j

Median 31.138 over the 676 units with a defined isolation distance (226 are NaN).

Screening task: Of the MTL cells, 138 (32.78%) qualified as concept cells (Fig. 4 shows examples).

Technical Validation, Proportion of selective cells, paragraph 2

Of 421 MTL screening units, 140 (33.25%) pass the permuted ANOVA and the permutation t-test, with the second test two-sided and one-sided alike (the same cells); both readings are within tolerance. Over the text's 449 the percent would be 31.18%, also within tolerance. Re-testing units near p = 0.05 with 20000 permutations gives 139 (33.02%). The code's parametric tests give 138 (32.78%), the paper's numbers exactly. Concept cells come from 16 of the 19 subjects with MTL units; subjects 13 and 14 hold 56 of the 140.

How the test differed from the paper’s method

• 2000 permutations, the number in the code's commented permutation alternative; the paper gives none. The p-value is (exceedances + 1) / (permutations + 1), as the code's permutationTest.m.
• The permutation t-test compares the difference of means with label permutations, as the code's permutationTest.m does; the Methods name a "permutation t-test" without saying which statistic is permuted.

In contrast, only 26 (5.35%) of cells in the MFC were concept cells, reproducing our earlier finding that concept cells are largely only found in the MTL.

Technical Validation, Proportion of selective cells, paragraph 2

Of 486 MFC screening units, 27 (5.56%) are concept cells under both readings, within tolerance of 26 and 5.35%, and far below the MTL's 33.25%. The count is near the upper bound of the tolerance (28.6); with three other seeds it is 29, 28 and 27, and the 29 (5.97%) would be outside. Re-testing the 142 MFC units with a p-value between 0.02 and 0.10 with 20000 permutations gives 26 (5.35%), within tolerance. The code's parametric tests give 26 (5.35%), the paper's numbers.

How the test differed from the paper’s method

• As for screening-concept-mtl (2000 permutations; difference of means permuted). The outcome at 2000 permutations depends on the seed for this claim, because the count lies near the tolerance bound; the 20000-permutation variant settles it within tolerance.

Shown are five ( a – e ) neurons classified as concept cells.

Figure 4 legend

All five units pass the concept cell test under both readings (permuted ANOVA p 0.0005 to 0.028, second test p at most 0.001) and under the code's tests. Four are in the amygdala and one in the hippocampus.

Working memory task: In the MTL, 94 (20.94%) neurons qualified as concept cells during the encoding period (Fig. 5 shows examples).

Technical Validation, Proportion of selective cells, paragraph 3

Of 449 MTL Sternberg units, 96 (21.38%) are concept cells with a two-sided second test and 100 (22.27%) with a one-sided one; both are within tolerance of 94 and 20.94%. The code's parametric tests give 94 (20.94%), the paper's numbers exactly. With three other seeds the counts are 95 to 97 (two-sided) and 99 to 100 (one-sided), all within tolerance. Over all areas the counts are 123, 128 and 119.

How the test differed from the paper’s method

• As for screening-concept-mtl (2000 permutations; difference of means permuted).

( a,b ) Example responses from 2 neurons classified as concept cells in the Sternberg task that also exhibited persistent activity during the maintenance period. Both cells were recorded in the amygdala.

Figure 5 legend

Both units (subject 7 unit 9, subject 16 unit 16, 1-based) are amygdala concept cells under both readings and the code's tests (all p = 0.0005), with the same preferred image (ID 5). Their mean maintenance rate is higher on trials that showed the preferred image (5.05 against 1.43 Hz, and 5.73 against 4.17 Hz).

In the MFC, 144 (31.79%) cells qualified as maintenance cells and 122 (26.93%) as probe cells.

Technical Validation, Proportion of selective cells, paragraph 3

Of 453 MFC Sternberg units, label shuffling gives 140 (30.91%, non-preferred as Encoding 1) and 139 (30.68%, preferred not shown), sign flipping 144 (31.79%) and 143 (31.57%); every concept cell set gives the same counts. All are within tolerance of 144 and 31.79%. Counting permuted values equal to the observed one as exceedances (on the same permutations) leaves label shuffling at 140 and 139 and gives 142 and 141 for sign flipping, all within tolerance. The code's paired t-tests give 144 (31.79%) and 143.

How the test differed from the paper’s method

• 2000 permutations, from the code's comment; the p-value as for screening-concept-mtl. Exceedances are counted with a strict inequality, as permutationTest.m does, which with discrete spike counts makes the tests slightly anti-conservative; counting ties is shown as a variant.

As a group, concept cells continued to fire during the maintenance period if their preferred stimulus was held in WM (4.76 ± 4.71 Hz vs. 2.67 ± 3.66 Hz; p = 1.38 × 10 −21 for pref. vs. non-pref. trials, paired one-sided t-test).

Technical Validation, Proportion of selective cells, paragraph 3

What the paper reports

• paired one-sided t-test, preferred against non-preferred trials, p = 1.4 × 10⁻²¹

In the maintenance window (0 to 2500 ms), under both cell readings, both trial readings and every concept cell set, concept cells fire more on preferred trials (every paired one-sided p below 1e-6), but the preferred means are 2.57 to 3.13 Hz against the paper's 4.76, and no reading puts the four values within tolerance. The measured values are reading (a)(i) with the two-sided concept test, MTL cells, preferred image in any encoding period (96 cells, 2.567 plus or minus 3.768 against 2.116 plus or minus 3.671 Hz, p = 3.9e-08). The variant in the Encoding 1 window (200 to 1000 ms) with preferred as the Encoding 1 image reproduces the paper's numbers exactly for the code's 119 parametric concept cells in all areas (4.764 plus or minus 4.709 against 2.673 plus or minus 3.656 Hz), and within tolerance for the permutation concept cells in all areas (123 cells, 4.631 against 2.588 Hz, p = 8.6e-22; 128 cells, 4.748 against 2.710 Hz, p = 1.3e-22). The paper's values therefore appear to be encoding responses. The direction the statement gives (more firing in maintenance when the preferred image is held) holds; the reported magnitudes do not. Leaving out the three trials with implausible timestamps (findings) and recomputing the concept cells (97 and 102 MTL cells, 94 parametric) gives maintenance-window preferred means of 2.57 to 3.10 Hz, no maintenance reading within tolerance, and 4.768 plus or minus 4.712 against 2.674 plus or minus 3.656 Hz for the parametric Encoding 1 variant, so these trials do not change the outcome.

How the test differed from the paper’s method

• The concept cells are those of sternberg-concept-mtl under its two readings; the code's parametric set is shown too.

In the MFC, 144 (31.79%) cells qualified as maintenance cells and 122 (26.93%) as probe cells.

Technical Validation, Proportion of selective cells, paragraph 3

All four readings were run on the 453 MFC units. Sign flipping gives 118 (26.05%, pooled encoding rate) and 111 (24.50%, Encoding 1), within tolerance; label shuffling gives 109 (24.06%) and 104 (22.96%), outside it. With three other seeds label shuffling with Encoding 1 stays outside (104 or 105), label shuffling with the pooled rate is just inside (110 or 111, bound 109.8) and sign flipping stays inside (110 to 124), so the readings disagree whatever the seed, and the criterion blocks the claim. Counting ties as exceedances on the main seed gives 108 and 103 for label shuffling and 118 and 111 for sign flipping, the same disagreement. The code's paired one-sided t-tests give 122 (26.93%), the paper's numbers. What is open is whether the Methods' "permutation t-tests" between probe and encoding or maintenance rates permute rates between the two periods as independent samples or keep each trial's pair (the code's own tests are paired t-tests, and its commented permutationTest.m is unpaired), and which encoding periods form the encoding rate. The passages are Methods, "Cell selection, probe cells" and the code's NWB_calcSelective_SB.m. The question is put to the paper's authors (findings).